Redefining Radiology - Artificial Intelligence Integration in Medical Imaging View PDF

Radiology has undergone a groundbreaking journey since its inception, marking a significant impact on modern medicine. From the discovery of X-rays to the subsequent incorporation of AI and ML, this multidimensional field continually evolves, transforming itself and the healthcare ecosystem it supports. This comprehensive examination examines the interaction between AI and ML in radiology, investigating their foundational principles, historical progression, practical implementations, inherent difficulties, and moral dilemmas. By enhancing comprehension of the contributions of AI and ML to radiology, this examination seeks to encourage insightful conversations among healthcare professionals, researchers, and policymakers, ultimately shaping the direction of the field and improving patient outcomes. The investigation delves into the fundamental concepts of AI and ML, their increasing influence in radiology, practical approaches to integration, and illustrative examples from various medical specialties [1, 2]. Additionally, it addresses challenges such as data accuracy, ethical considerations, and contemplates potential future paths in AIdriven radiology.

The role of radiology in modern medicine

Radiology, the medical field focused on using different imaging techniques to diagnose and treat illnesses, has become a fundamental part of present-day healthcare, playing a crucial role in clinical practice. It goes beyond simply identifying diseases and includes providing guidance for treatment and ongoing disease management. Proficiency in diagnostic techniques like CT (computed tomography) scans, MRIs (magnetic resonance imaging), PET (positron emission tomography) scans, ultrasounds, and X-rays are used to guide immediate medical interventions, monitor treatment progress, and visually document a patient’s health status. The detailed understanding of anatomical, physiological, and molecular disease processes that medical imaging provides has a significant impact on patient care, allowing for personalized treatments that improve outcomes and minimize adverse effects [3]. Radiology is an essential component of interdisciplinary medical teams, with radiologists providing accurate and timely imaging reports that enhance communication between specialists and influence important decisions, ultimately contributing to a holistic approach to patient-centered healthcare. As valued consultants, radiologists offer valuable insights into the selection and interpretation of appropriate imaging studies, while also playing a crucial role in ensuring radiation safety and dose management [4, 5]. Their expertise helps to paint a clearer clinical picture, providing insights that can greatly influence patient care.

A brief history from Wilhelm Roentgen’s groundbreaking discovery to magnetic fields

Wilhelm Roentgen’s discovery of X-ray technology in 1895 to current advanced methods, the evolution of modern medical imaging showcases the unyielding quest for scientific progress and its profound influence on radiology. Roentgen’s unparalleled X-ray breakthrough provided a non-intrusive insight into the human body, establishing the basis for contemporary imaging. Despite its initial shortcomings in depicting 2D images and distinguishing soft tissues, this underlying principle set the stage for more intricate, non-intrusive imaging techniques [6].

Sir Godfrey Hounsfield and Allan Cormack’s introduction of CT in 1973 marked a significant breakthrough, surpassing the limitations of 2D imaging by presenting a 3D format. A CT scan works by rotating X-ray sources and detectors around a patient’s body synchronized to differential absorption. By combining this with advanced computational algorithms, 3D volumetric data can be reconstructed from 2D images. Ultrasound imaging, which emerged in the 20th century, deviated from radiation-based technologies by utilizing high-frequency sound waves to generate real-time images of internal body structures. Its non-ionizing radiation properties, real-time imaging capabilities, and cost-effectiveness have made it widely applicable in various clinical fields, including obstetrics, gynecology, cardiology, and emergency medicine. Particularly in emergency and critical care medicine, pointof- care ultrasound (POCUS) has played a vital role in facilitating rapid bedside assessments and expediting clinical decision-making [7]. In the 1970s, MRI was developed by Paul Lauterbur and Sir Peter Mansfield. This technology utilizes a powerful magnetic field and radio waves to produce highly detailed images of the body, especially of soft tissue structures. The non-ionizing nature of MRI, combined with its exceptional ability to differentiate soft tissues, has revolutionized medical imaging [2]. Manipulation of RF pulse sequence timing in MRI has further improved its diagnostic usefulness, allowing for the acquisition of various image types and the identification of distinct tissues and pathologies (Figure 1).

An evolution in radiology- transition from film to function

In addition to the disruptive advances in imaging methods, there was a significant change in the late 20th century: the shift from filmbased to digital radiography and the introduction of Picture Archiving and Communication Systems. This transition greatly improved the efficiency of acquiring, storing, and retrieving images, while also facilitating the seamless sharing and transfer of images within and between healthcare institutions [8]. The advancements in medical imaging technology did not cease with these developments. Functional imaging techniques, such as PET, renowned for its utilization of radiolabeled biochemical substances, and Single-Photon Emission Computed Tomography (SPECT), which employs gamma-emitting radionuclides to track biological processes, have shed light on metabolic and biological functions. These techniques provide a glimpse into cellular activity and offer invaluable insights into the functional condition of organs. Moreover, the progress in 3D imaging has revolutionized medical imaging by enabling a more precise understanding of spatial connections within the body. Consequently, this has enhanced diagnostic precision and surgical preparation considerably [9, 10]. The subsequent development of 4D imaging extended the limits by incorporating the temporal aspect, enabling the real-time observation of biological processes. The merging of functional and structural imaging led to the emergence of hybrid imaging technologies like PET/CT and SPECT/CT. These methods merge the benefits of both approaches, delivering extensive diagnostic data. For example, PET/ CT combines the metabolic insights of PET with the precise anatomical context provided by CT, greatly enhancing the precision of lesion identification and description. Interventional radiology, using imaging for minimally invasive procedures, has transformed healthcare. Realtime visualization improves precision, enhances patient outcomes, and shortens recovery periods [2]. Image-guided biopsies are a safer and less invasive alternative to surgical biopsies with fewer complications and shorter hospital stays.

New frontiers in future

The integration of virtual reality/augmented reality (VR/AR) and AI is set to revolutionize the field of radiology, ushering in a new era in medical imaging. VR/AR technologies, which originated in the gaming and entertainment sectors, are gradually making their way into radiology, creating an immersive environment that is beneficial for both radiology training and clinical practice [11]. In the latter, these technologies have the potential to enhance the visualization of imaging data, thereby improving diagnosis and treatment planning. Meanwhile, AI, specifically machine learning, is making significant advancements in radiology, enhancing image analysis, and reducing diagnostic errors. Through the use of AI algorithms, data can be processed and interpreted, allowing for tasks that simulate or even surpass human cognitive abilities. Through the utilization of labeled examples, ML has the ability to extract intricate and advanced information from datasets that do not have labels. By combining AI with VR and AR technologies, there is a significant potential for enhancing the efficiency of radiology, improving the accuracy of diagnostics, and exponentially transforming treatment planning. Over the past twenty years, the field of radiology has refined computer-aided diagnosis (CAD) tools that are based on ML. These tools are poised to bring about an integrated diagnostic service that incorporates radiology, pathology, and genomics data, thereby enhancing the performance of CAD and improving the productivity of radiology services through AI-assisted workflows. Nevertheless, the integration of AI and VR/AR in radiology faces technical challenges, particularly in terms of incorporating AI-generated results into existing workflows. However, a proposed roadmap suggests the integration of AI-based image analysis algorithms, which would include a feedback loop system between radiologists and AI, enabling continuous improvement. An instance where AI and radiologists worked together was showcased through a study that showed enhanced identification of brain metastases. The ethical considerations surrounding the use of AI in radiology require thorough examination. While AI shows promise in radiology, ethical concerns arise, particularly regarding biases and the lack of transparency in AI decision-making [12-16]. It is crucial to advocate for ethical AI to ensure that discriminatory effects and injustices are avoided. Future advancements in AI in radiology should integrate perspectives from the field of social science.

The Progression of AI and ML

The purpose of this section is to provide an overview of the development of AI and ML, as well as explain their distinct yet interconnected terminologies and historical trajectory. As well as shedding light on the key machine learning algorithms and techniques that have shaped our technological landscape, it emphasizes how indelible their imprint is still present today [2].

Breakthroughs of AI

Historically, the development of AI, characterized by folklore and tales of artificial beings, has been rooted in the philosophy of human cognition as a mechanistic process, dating back to antiquity. In the 1940s, programmable digital computers were developed, leading to the establishment of the field at the 1956 Dartmouth Conference, marking the beginning of modern AI concepts [17].

In the 1970s, MYCIN, an expert system that totally changed the game. It was created by Buchanan and Shortliffe, and the idea was to mimic human expertise using knowledge bases and inference engines. This was a huge deal because it made AI a big player in medical diagnosis and decision-making. Then, machine learning algorithms came along and shook things up even more. Decision trees in 1986, support vector machines in 1995, and neural networks in 1986 expanded the AI world in healthcare. These algorithms analyze tons of data and opened up a whole new era of pattern recognition and predictive modeling in healthcare [18,19]. The turn of the century witnessed a shift with the arrival of deep learning (DL) approaches, specifically convolutional neural networks (CNNs). CNNs, modeled after the structure and function of the human brain, surpassed previous methodologies in image recognition tasks. They played a huge role in improving medical image classification, segmentation, and detection, all because they can learn from a lot of labeled data. The fast progress of AI in recent years is thanks to two main factors: the rise of big data and improvements in computational power. We now have a lot of electronic health records (EHRs), medical image collections, and datasets with annotations that we can use for training. And with advancements in hardware, like graphical processing units and distributed computing, we can run AI algorithms that require a lot of computational power much faster [20]. In the last few years, the rapid progress of AI language models, such as GPT-4, has led to diverse applications and implications across a wide range of industries, including the healthcare and medical imaging industry [2]. These models, capable of producing text that resembles human language and facilitating communication, have raised significant concerns. The undeniable potential of these models to make substantial contributions to medical research and patient care is tempered by experts expressing reservations about their limitations and the potential for inadvertently perpetuating inequalities or disseminating inaccurate information. Hence, it is essential to adopt robust strategies to responsibly manage these risks, which involve enhancing transparency regarding potential harms, enabling early identification of issues, and implementing regulatory measures and peer evaluations. Through these strategies, the goal is to ensure the optimal and ethical utilization of AI technologies, including language models, thereby positively impacting medical imaging and healthcare outcomes. Recognizing these milestones enables us to acknowledge the evolution of AI in healthcare, particularly in the field of medical imaging [21]. Currently, as a result of the interaction between rulesbased systems, traditional machine learning algorithms, as well as the transformative impact of DL techniques, we have established the foundation for the current state-of-the-art AI applications in radiology and other medical specialties (Figure 2).

Decoding AI, ML, and DL

The landscape of computational intelligence is a vast system, comprised of separate yet interconnected systems, with each sector are two main types of machine learning: supervised and unsupervised learning [27]. Supervised learning, which dominates the ML landscape, utilizes established examples or training data, comprising input-output pairs. The goal is to create a function that can accurately predict or classify unfamiliar data by mapping input data to corresponding outputs. Key algorithms in supervised learning include linear regression, logistic regression, and decision trees. On the other hand, unsupervised learning explores the data space independently to uncover inherent patterns, structures, or relationships without predetermined outputs [28, 29]. Its focus lies in identifying intrinsic data structures, thereby providing insights that have the potential to address complex problems. Prominent unsupervised learning algorithms encompass clustering techniques like k-means and hierarchical clustering, as well as dimensionality reduction methods such as principal component analysis. ANNs mimic the operational framework of the human brain, carrying out intricate tasks through a network of interconnected artificial neurons organized in layers. The backpropagation algorithm, a crucial component of ANN operations, demonstrates a strong ability to handle faults, ensuring the system’s functionality even in the face of occasional neuronal failures [30]. ANNs enable the extraction of features and the identification of complex patterns, which are essential for ML. This, in turn, improves data representation and class differentiation by assisting with the preprocessing of raw data for feature extraction or selection. Advances in ANNs have led to the development of complex structures such as DL models that consist of multiple layers of neurons. Notably, CNNs utilize convolution instead of standard matrix multiplication in specific layers. Designed for processing pixel data, CNNs excel in tasks that involve pattern recognition in images, audio, or text. They significantly contribute to computer vision and Natural Language Processing (NLP) by simplifying complex patterns into abstract representations through layers of features [2].

Integrating AI into Medical Imaging

This section explores the significant role of AI in the field of medical imaging, highlighting the groundbreaking changes it triggers in radiology. The versatile potential of AI, ranging from transforming image acquisition to reshaping radiological analyses, and from optimizing reporting to creating individualized medical stories, places AI at the center of the ongoing healthcare revolution [31]. Beyond radiology, this transformation extends to other areas of healthcare, including pathology, cardiology, genomics, drug discovery, and healthcare delivery, where the impactful advancements of AI are increasingly acknowledged. Wrapping up this examination is the emerging concept of AI-enabled personalized medicine, emphasizing a more proactive, patient-centered, and comprehensive patient care approach (Figure 3) [32].

A paradigm shift in radiology

AI has sparked a profound transformation in the realm of radiology, reshaping conventional processes and enhancing the role of radiologists. In the domain of image capture, AI enhances scanning techniques, optimizes image accuracy, and promotes advanced image reconstruction in MRI, CT, and PET procedures. Of particular note, deep learning expedites MRI scans, promoting efficiency and quality, with similar advancements observed in CT and PET image reconstruction [33-36]. AI significantly streamlines the process of radiologist analysis on chest X-rays, as demonstrated by a study where an AI system reduced interpretation delivery time from 11.2 days to just 2.7 days, underscoring the effectiveness of automated triage systems in streamlining healthcare workflows and improving patient care standards.

Radiology, being one of the early adopters of digital technology in healthcare, has effectively utilized ML in CAD tools for more than twenty years, demonstrating strong performance in both sensitivity and specificity. Despite facing various challenges, the integration of AI is seen as a crucial solution to overcome these obstacles, improving CAD performance, making radiology services more efficient, and promoting the development of integrated diagnostic services. The incorporation of AI in radiology reporting has resulted in structured and annotated data, ensuring consistent reports, and facilitating patient history tracking. These advanced tools generate comprehensive task lists that include relevant information from the patient’s history, aiming to enhance the accessibility and integration of reports into care pathways. In addition to revolutionizing reporting and imaging procedures, these cutting-edge systems also play a vital role in maintaining communication between providers and ensuring quality patient care by validating correlations between imaging diagnoses, radiological reports, and treatment plans, as well as identifying any discrepancies. Furthermore, AI optimizes the allocation of personnel and scanner usage, while also reducing radiation exposure, thereby increasing efficiency and the overall quality of care. With its wide-ranging capabilities, AI is reshaping the field of radiology and solidifying its essential role within the discipline.

Applications of AI in healthcare

The field of healthcare has been significantly impacted by AI, showcasing its ability to improve various aspects of clinical practice beyond just radiology. It has proven its worth in diagnostics, genomics, drug discovery, and optimizing healthcare delivery. Pathology, in particular, has seen tremendous progress with the implementation of AI algorithms for tissue analysis. This has greatly enhanced the accuracy and speed of diagnoses. Furthermore, automated image analysis tools have allowed pathologists to examine tissues at a microscopic level, identifying subtle histopathological characteristics that are often missed by human observers. AI has also played a crucial role in advancing digital pathology by transforming traditional glass slides into digital scans [37]. This enables remote diagnostics and collaborative work, which are essential in the digital era of telemedicine. AI has shown great potential in the field of cardiology as well, especially in the interpretation of electrocardiograms and echocardiograms. Advanced ML algorithms identify intricate cardiac patterns and irregularities, precisely forecasting conditions such as atrial fibrillation and heart attack. The exponential growth in echocardiography is apparent, with automated algorithms assisting in the interpretation of cardiac structure parameters, reducing differences in observation between individuals, and improving diagnostic accuracy. The intricate nature of genomics makes it a perfect candidate for AI intervention [38]. DL techniques decode genomic data, assisting in the identification of genetic variations associated with disease vulnerability, and creating opportunities for customized treatment approaches based on individual genetic profiles. AI has proven to be invaluable in the realm of drug discovery by expediting the search for powerful therapeutic compounds and subsequently expediting the drug development process. For instance, AI can anticipate the pharmacokinetic and pharmacodynamic properties of novel compounds, identify potential drug targets, and simulate clinical trials, significantly reducing the time and costs associated with drug development. Additionally, AI’s ability to optimize healthcare delivery is remarkable [2]. AI-driven predictive analytics can improve hospital workflows by accurately predicting patient admission rates and optimizing the allocation of resources. AI applications for cost reduction have also emerged, with ML algorithms identifying inefficiencies in healthcare systems, thus facilitating cost-effective care. The ultimate goal of these AI applications is to enhance patient outcomes by streamlining diagnostic procedures and personalizing treatment plans [39-41].

Applications of AI in radiology

This portion takes a critical look at the practical uses of AI in the field of radiology, delving into the innovative methods it brings to imaging techniques, diagnosis, and patient care. By exploring AIdriven approaches like DL and CNNs, the portion highlights how AI is revolutionizing the process of image segmentation and classification, as well as diagnostics [42, 43]. Additionally, it investigates the predictive capabilities of radiomics and the potential of AI in optimizing workflows. Throughout the discussion, the portion also addresses the inherent difficulties and obstacles in integrating AI in radiology, emphasizing the importance of interpretability, validation, standardization, and the preservation of the human element in healthcare [44].

Image classification and segmentation

DL has sparked a notable change in the field of radiology, particularly in the areas of image partitioning and categorization, where significant progress has been achieved. The advancements brought about by these AI-centric methods have increased the accuracy and speed of diagnoses, thereby improving the abilities of radiologists and enhancing patient care. However, the integration of AI poses various challenges that must be overcome to ensure its successful implementation and use in radiology. CNNs, with their inherent ability to learn complex patterns through backpropagation, have emerged as powerful tools for computational visual tasks, including various applications in radiology [45]. Their unique architectural layers, which encompass calculations in convolutional layers and predictions in fully connected layers, combine to create an efficient system for detecting objects. These networks have exhibited impressive proficiency in object detection, thanks to their integrated capabilities in feature extraction, semantic segmentation, and handling multi-scale features. The efficiency and effectiveness of CNNs can be further enhanced through the utilization of transfer learning, which allows for the reuse of existing models. This approach improves accuracy, enables efficient training with limited datasets, and reduces the need for labor-intensive and error-prone manual segmentation. An example of CNNs’ potential is the segmentation of lung nodules from CT scans using AI, which has demonstrated excellent performance in the early detection and treatment of lung cancer. It achieved an AUROC of 94.4%, outperforming six radiologists in the task. CNNs have played a crucial role in the segmentation of brain tumors from MRI scans and the examination of retinal images for early indications of diabetic retinopathy, further highlighting the wide-ranging applicability and adaptability of AI in the field of medical imaging [46]. The use of CNNs in image classification, another important AI application in radiology, enables the differentiation between normal and abnormal findings.

AI models have been developed to distinguish between benign and malignant tumors in mammography, achieving performance metrics that are comparable to those of human radiologists. This enables the early detection of breast cancer. However, there are various challenges to integrating AI into radiology. Developing robust and dependable AI models requires access to extensive datasets and thorough validation, which can be difficult due to privacy concerns and the diverse nature of medical imaging data. Additionally, ensuring the interpretability and explainability of DL models, which are often seen as “black boxes,” is crucial for gaining trust and acceptance from both radiologists and patients. Lastly, incorporating AI into existing clinical workflows and promoting a harmonious collaboration between humans and AI are vital for fully realizing the potential of AI in radiology and translating these technological advancements into tangible improvements in patient care [47, 48].

Prognostics with radiomics and predictive analytics

Radiomics is a burgeoning area in the medical field that revolves around extracting complex data from radiological images. It holds immense promise for medical diagnostics, prognosis, and assessing how diseases respond to treatment. Nevertheless, radiomics encounters obstacles such as the need for standardization and validation to guarantee consistent and dependable results. The primary advantage of radiomics lies in its capacity to supplement traditional clinical practices with precise and quantitative information, thereby transforming the decision-making processes in medicine. The exponential growth of medical imaging data has created an ideal environment for the utilization of ML and data-driven science. Radiomics-based decision support systems for accurate diagnosis and treatment are on the verge of becoming indispensable tools in modern medicine [51]. However, it is important to acknowledge that the journey of radiomics towards full clinical applicability is not devoid of challenges. The field currently grapples with the essential requirements of standardization and validation to ensure trustworthy and reproducible outcomes. In this context, the emergence of AI offers promising opportunities to overcome these challenges and unlock the complete potential of radiomics. By utilizing AI-driven analytics, accurate predictions regarding disease progression, treatment response, and patient survival can be made by modeling complex data sets. This advancement provides clinicians with an unprecedented amount of information that goes beyond what human perception alone can achieve. In the field of oncology, radiomics has played a vital role in identifying different molecular traits and the spread of cancer to lymph nodes, as well as evaluating treatment response and predicting disease survival. It is important to recognize that the integration of AI into radiomics is still in its early stages. To fully utilize the potential of AI in medical imaging, there needs to be a focused effort on research and development [52-53]. This includes promoting large-scale data sharing, establishing standardized protocols for data collection, defining clear evaluation criteria, and implementing robust reporting guidelines. These components are essential for the growth and widespread adoption of radiomics as a field, ushering in a new era of precision medicine.

Optimization of workflow using AI

AI is gaining momentum in the field of radiology, with the goal of optimizing workflows and improving the effectiveness of noninterpretive tasks. When combined with NLP, AI has the ability to automate the prioritization of imaging studies, placing urgent cases at the top of the list by extracting and analyzing key patient data from EHRs. This streamlines the process of triaging patients, generating radiology reports, and managing follow-ups for incidental findings. The implementation of AI greatly enhances the radiology process by automating triage and enhancing report generation [2]. It efficiently sorts and prioritizes radiological studies such as CT scans and MRIs based on urgency, highlighting critical cases that require immediate attention. This automated system helps consistently detect serious conditions like stroke, hemorrhage, and cancer, thereby reducing errors. The use of AI, particularly NLP, in non-interpretive tasks helps alleviate the monotonous aspects of the workflow, potentially mitigating radiologist burnout. Additionally, AI improves the creation and interpretation of radiology reports. DL algorithms address the shortcomings of traditional reporting, including errors caused by fatigue or inconsistencies due to varying levels of expertise [54]. These algorithms detect and characterize findings to enhance consistency, standardize report creation, and minimize errors. This additional layer of analysis streamlines the workflow and enhances the clarity of reports, making a significant contribution to the quality of radiology services.

The integration of AI surpasses the boundaries of purely diagnostic capabilities, greatly enhancing interdisciplinary collaboration and communication between patients and radiologists. AI platforms act as a crucial channel that fosters a shared understanding of imaging results among diverse healthcare professionals by simplifying complex medical terminologies for patients. This transparency helps establish a strong relationship between the patient and the radiologist, while also promoting greater patient involvement in their own healthcare. While progress is being made integrating AI into radiology, it is important to note that as of 2021, only 30% of radiologists reported using clinical AI, with more than 70% expressing hesitation to invest in AI. Many views AI as providing minimal advantages, suggesting that the field of radiology is currently in a phase of disappointment in the adoption of AI [56- 59]. This disappointment stems from factors such as doubt about AI performance and applicability, a perceived lack of necessity, inefficient workflows for AI utilization, and a lack of scalable infrastructure to support AI. In order to transition into a phase of enlightenment, the field must establish infrastructure that enables optimal AI functionality, which includes redefining and disrupting existing systems such as image management and Picture Archiving and Communication Systems for intelligent workflow coordination [65]. Despite the potential of AI, the importance of preserving the human element in patient care cannot be overstated. This emphasizes that while AI can enhance the work of radiologists, it cannot replace the nuanced judgment and empathetic communication that are essential in-patient care.

AI across medical specialties

Neuroradiology

ML, especially supervised techniques, and DL have become essential in handling complex data in the field of neuroradiology. This state-of-the-art technology has enabled the early identification of various types of strokes, as shown in multiple studies. CNNs exhibit exceptional proficiency in tasks such as detecting infarcts or hemorrhages, segmenting data, classifying strokes, and identifying occlusions in large blood vessels. The utilization of CNNs in these domains has greatly impacted stroke treatment approaches, as illustrated by the research conducted by Soun et al.

AI enhances clinical decision-making in situations characterized by significant variation between raters. Its applications range from categorizing different types of strokes and identifying bleeding in the brain to segmenting images and identifying blockages in large blood vessels. This advancement offers clear advantages for facilities that handle a small number of stroke patients or serve as regional centers. A growing body of research highlights the potential of AI in supporting decisions regarding the use of thrombolysis and thrombectomy [60,61]. For example, an AI model developed by researchers was able to accurately identify blockages in large blood vessels using CT scans, demonstrating a high level of accuracy in determining which patients would benefit from timely intervention. Zhu et al. utilized AI algorithms to predict how patients with acute ischemic stroke would respond to thrombolysis, combining imaging features with clinical data to assist clinicians in developing effective treatment plans. AI also plays a crucial role in the early detection of neurodegenerative disorders, particularly Alzheimer’s and Parkinson’s diseases. Sophisticated AI algorithms have been created to analyze MRI images and identify specific biomarkers or patterns associated with these conditions. AI’s ability to detect subtle changes in brain structure or function at the voxel level makes it an efficient tool for diagnosing these diseases and providing objective, quantitative assessments [62-64]. Additionally, AI has shown promise in predicting outcomes for brain and spine surgeries. By analyzing preoperative imaging data, AI models can provide prognostications about surgical outcomes, including the likelihood of complications and the extent of functional improvement.

Oncological imaging

Progress in high-performance computing has led to significant advancements in oncology, specifically in cancer imaging, thanks to the development of AI and ML technologies. Through the integration of multi-omics data and the use of DL strategies, precision oncology has been able to make great strides in cancer diagnosis, prognosis, and treatment. The digital nature of oncological imaging makes it wellsuited for AI and ML applications, as the entire imaging process, from acquisition to interpretation, reporting, and communication, takes place within the digital space. This allows for efficient data capture and analysis using AI and ML. Consequently, these technologies are now being actively explored and implemented in cancer imaging, which is a major component of the workload in many healthcare facilities. AI is increasingly being used in tumor detection and classification, particularly in the diagnosis of breast, lung, and prostate cancers. AIbased devices are already being utilized in clinical practice. Studies have shown that deep learning models and CNNs can accurately classify lung nodules on CT scans and differentiate subtypes of renal cell carcinoma on MRI, often matching or even surpassing the expertise of experienced radiologists. AI algorithms provide an objective and consistent way to assess changes in tumor size or metabolic activity, automating measurements that were previously time-consuming and prone to variability between observers, as seen in the Response Evaluation Criteria in Solid Tumors. AI achieves this by utilizing radiomic features—complex data obtained from radiological images— to construct mathematical models proficient at detecting subtle changes suggestive of treatment response [65]. Additionally, AI plays a crucial role in monitoring treatment response by quantifying tumor changes through detailed analysis of medical image subunits (pixels/voxels). These smaller components can be examined by computers to reveal objective mathematical characteristics associated with disease behavior or outcomes. AI also obtains valuable predictive insights by analyzing radiomic signatures, such as texture analysis, to forecast survival rates in lung cancer patients based on pre-treatment CT images, while radiomic features derived from MRI scans have demonstrated a correlation with the risk of recurrence in patients with glioblastoma. Consequently, the incorporation of AI in radiology enables efficient and accurate tracking of tumor progression, significantly improving overall treatment evaluation and patient care. Traditional methods of monitoring radiation therapy response, which rely on manual assessment of changes in tumor size and characteristics, are often subjective and may fail to detect subtle indications of treatment effectiveness [66]. AI, specifically CNNs, offer an objective approach to assess treatment response, utilizing extensive datasets of annotated imaging scans to precisely identify and delineate tumors. This automation streamlines the planning process, potentially enhancing treatment outcomes through precise radiation dosage. The effectiveness of this approach is frequently measured by the Sørensen– Dice coefficient, facilitating early and precise assessment of therapy efficacy and prompt treatment adjustments, if necessary (Figure 4).

Cardiovascular imaging

In recent years, there has been significant progress in cardiovascular imaging thanks to AI. This progress has allowed for better detection and measurement of heart diseases, thorough analysis of vascular abnormalities, and integration of various imaging data. AI algorithms efficiently interpret complex imaging data, identifying early stages of cardiac diseases like coronary artery disease and congestive heart failure using modalities such as cardiac CT, MRI, or echocardiography. For instance, ML models and CNNs have shown the ability to automatically detect calcification in coronary arteries and segment the left ventricular myocardium, respectively [67-70]. These automated processes have shown strong correlation with manual analyses. Furthermore, AI technology has improved the assessment of the left ventricle in echocardiographic diagnosis by automating tasks that were traditionally reliant on visual observation and manual boundary tracing. This includes measuring the left ventricular ejection fraction using the Simpson method. These advancements aim to reduce reliance on physician experience, thus enhancing the repeatability and accuracy of evaluations. In addition to cardiac conditions, AI is also valuable in analyzing vascular abnormalities such as aortic aneurysms and peripheral artery disease. This technology facilitates early intervention and potentially improves patient outcomes. CNNs have proven to be effective in evaluating abdominal aortic aneurysms from CT images, displaying high accuracy in detecting and sizing these life-threatening conditions. Moreover, AI-assisted standard section recognition has significantly reduced evaluation time, improved detection capabilities, and enhanced the accuracy of novice practitioners. These benefits are particularly valuable in settings with limited resources for training echocardiography physicians.

The fusion of data from multiple imaging techniques, including CT, MRI, and echocardiography, marks a significant advancement in AIbased cardiovascular imaging [71]. This comprehensive data integration enables a comprehensive analysis of the structure and function of the heart, which is vital for intricate evaluations like identifying ischemia or preparing for medical procedures. An example of this integration is the utilization of machine learning algorithms to combine perfusion data from MRI and coronary anatomy from CT, resulting in the creation of advanced 3D models of the heart. This not only enhances the detection of cardiac ischemia but also aids in accurate procedural planning.

The domain of gastrointestinal imaging has experienced unparalleled progress thanks to AI. AI has played a crucial role in enhancing the identification, diagnosis, and staging of liver and pancreatic diseases [72]. The creation of various predictive models based on AI has expanded the range of diagnoses, including gastrointestinal and inflammatory diseases, non-malignant conditions, and the detection of bowel bleeding using state-of-the-art technologies like wireless capsule endoscopy. Moreover, AI has proven to be invaluable in the identification of hepatic-associated fibrosis by utilizing EHRs to gain meaningful insights into patients’ health data and medical history [73]. Additionally, the fusion of AI with endoscopic ultrasound technology has significantly improved the speed and accuracy of diagnosing pancreatic carcinoma, thereby enhancing patient management strategies.

AI has revolutionized the way liver and pancreatic diseases are diagnosed in the fields of hepatology and pancreatology. The integration of AI with various imaging techniques has led to a significant shift in the diagnostic approach. Ultrasound, endoscopic ultrasonography, CT, MR, and PET/CT have all benefited from AI’s advancements. Additionally, AI plays a crucial role in helping doctors choose the most suitable diagnostic test for each patient based on their unique medical profile. Furthermore, AI is invaluable in improving image quality, speeding up image acquisition, and predicting patient prognosis and treatment response (Figure 5) [74].

AI has revolutionized the field of abdominal and pelvic imaging by providing precise and consistent diagnostic results. This technology enables automated or semi-automated identification and alignment of the liver and pancreatic glands, as well as their associated abnormalities, leading to improved accuracy in diagnosis and treatment. The integration of radiomics introduces new quantitative measures in radiology reports, which enhance the identification and characterization of localized lesions and widespread disorders in the liver and pancreas, ultimately leading to better clinical outcomes [2].

In the field of nephrology, advanced AI applications exhibit great potential in predicting the development of acute kidney injury even before significant biochemical changes occur. This early detection allows for timely intervention to prevent the progression of the disease [75]. Furthermore, AI’s capability to identify modifiable risk factors for the progression of chronic kidney disease provides valuable insights for preventive care.

As a result, AI models have demonstrated proficiency in interpreting imaging studies that match or even surpass human accuracy in the field of renal tumor detection. After renal transplantation, this impressive feat could enhance prognostication and decision-making processes [76]. As renal tumors are more precisely detected and diagnosed, treatment strategies become more effective, potentially leading to better patient outcomes.

Limitations and Future Directions

The rise of AI in the healthcare industry, specifically in diagnostic radiology, has opened up new possibilities to improve the level and effectiveness of patient treatment. However, along with this rapid development, there are numerous obstacles that need to be overcome. These challenges include the need for sufficient and accurate data, the complex nature of AI algorithms, integration into medical practices, and ethical concerns [77]. This section will explore these issues in detail and suggest potential solutions to promote the integration and responsible application of AI in radiology, taking into account technical, infrastructural, regulatory, and human factors.

The “black box” problem and data quality

The effectiveness of AI algorithms, which essentially mirror reality through mathematical calculations, relies not just on the quality and accuracy of their training data, but also on the appropriate interpretation of medical images based on comprehensive datasets that accurately capture the diversity of patient demographics, including age, gender, ethnicity, and disease progression [78].

However, the creation of such datasets is often hindered by biases caused by the limited inclusion of certain demographic groups or specific clinical contexts. To address this issue, techniques such as data augmentation, oversampling, and under sampling are commonly used to mitigate data scarcity, ensuring that the training dataset is diverse and represents a balanced view of the subject matter.

Recognizing and addressing potential risks related to biased or non-representative data is crucial. Failure to manage these risks adequately can unintentionally perpetuate health disparities and result in AI models that perform poorly for certain patient groups. The issue of “black box” in AI, which refers to the lack of transparency in AI models, makes it difficult to detect errors and identify biases, thereby negatively impacting underrepresented communities and the effectiveness of clinical applications [79-81]. Overcoming these challenges requires a collaborative effort to diversify data collection, address bias in AI system design, analyze performance based on population subgroups, and utilize representative samples for clinical validation.

In addition to these approaches, explainable AI is an emerging field that aims to demystify AI decision-making processes by improving the reliability of inferences and enhancing transparency and interpretability. Explainable AI techniques, such as saliency maps, feature importance, and surrogate models, help visualize and explain the reasoning behind AI models’ decisions, making them understandable to both experts and non-experts.

The inclusion of AI in clinical radiology has received mixed responses, requiring a clear plan to address the various challenges that emerge. Key among these is the need for strong hardware and dependable software, crucial for managing the large amount of data produced by medical imaging systems [83]. AI-powered applications can tap into the untapped potential of around 97% of unused hospital data, greatly enhancing the prediction of disease progression and the adjustment of treatment plans. 

Implementing AI effectively necessitates high-performance hardware capable of carrying out complex calculations in real time.This implies a significant investment in powerful and reliable hardware infrastructure to avoid any negative impact on patient care caused by system failures. Additionally, the software must be resilient, userfriendly, and designed to seamlessly integrate into existing radiological systems [84]. This requires close collaboration between AI developers, radiologists, and other healthcare professionals. AI also offers the promise of reducing the administrative burden faced by radiologists, who currently spend approximately 16.6% of their working hours (roughly nine hours per week) on administrative tasks.

It is equally crucial to overcome the challenges of obtaining regulatory approval and maintaining ongoing monitoring. AI tools must undergo thorough validation to prove their safety and effectiveness before receiving regulatory approval. After approval, it is essential to continually monitor and evaluate their performance to ensure consistent improvement and reliable outcomes. Ultimately, the successful integration of AI in radiology depends on the acceptance and adoption by end-users, particularly radiologists [85]. This emphasizes the significance of offering adequate training and support to enable radiologists to utilize AI tools effectively and seamlessly incorporate them into their daily routine.

Ethics: conundrums and dilemmas

AI integration in radiology raises several ethical concerns, including issues with data privacy and security, patient confidentiality, informed consent, the risks of misdiagnosis, and the need to maintain human involvement in patient care.

The convergence of data privacy and security with AI revolves around how patient data is governed, especially in public-private partnerships with large tech companies. An examination of commercial healthcare AI has revealed that a significant portion of these technologies is controlled by private entities, causing concerns about potential misuse of data [86]. The DeepMind incident serves as a prime example, as it involved the transfer of patient data from the UK to the US without explicit patient consent. These incidents highlight the necessity for stricter regulatory oversight to ensure that patient data remains within its original jurisdiction and is protected against unauthorized access.

Patient privacy and consent, closely related to data protection, necessitate patients to have control over their data, understanding how it is used, the potential risks involved, and the benefits it may bring. This becomes particularly important in the context of AI-driven healthcare, where the opaqueness of learning algorithms can obscure the decision-making processes. Therefore, it is crucial to establish necessary measures and transparent procedures to ensure privacy and protect patient autonomy.

Moreover, the ethical concerns surrounding AI-assisted radiology include the possibility of misdiagnosis, liability, and accountability. While AI shows promise in providing diagnostic capabilities, it is also prone to errors and biases that can result in incorrect diagnoses and harm to patients [87, 88]. Addressing this issue requires the formulation and implementation of clear guidelines and policies for the use of AI in medical settings, with a focus on holding decision-makers accountable and defining the responsibilities of healthcare professionals and AI systems.

Furthermore, the ethical issues surrounding AI-supported radiology encompass the potential for misdiagnosis, legal responsibility, and answerability. Despite the potential for AI to offer diagnostic capabilities, it is also susceptible to mistakes and prejudices that can lead to incorrect diagnoses and harm to patients. Addressing this matter necessitates the development and implementation of clear guidelines and policies for the utilization of AI in medical environments, with an emphasis on holding decision-makers accountable and defining the obligations of healthcare professionals and AI systems.

Having said that, the danger of excessive reliance on AI underscores the significance of upholding the human factor in patient care. While AI can enhance healthcare delivery and complement the expertise of healthcare professionals, it should not overshadow the invaluable knowledge and nuanced decision-making inherent in the practice of medicine, particularly considering the intricacy and variability of individual cases [89]. As the role of AI in radiology continues to evolve, these moral considerations must remain prominent in discussions, policy development, and research, ensuring the responsible and fair application of this transformative technology in healthcare.

Collaboration between radiologists and AI developers

The convergence of AI and radiology demands a collaborative partnership between radiologists and AI experts to establish cuttingedge solutions that promote medical progress and enhance patient wellbeing. This section explores the amalgamation of these distinct domains and the significance of interdisciplinary cooperation, its advantages, and approaches to bridging the divide between academia and industry.

Radiologists contribute a profound comprehension of clinical requirements, disease mechanisms, and nuances in interpreting medical images, while AI experts possess the technical proficiency to devise, implement, and optimize ML algorithms. This diverse range of expertise is not only complementary but also indispensable for the development of AI tools that are both feasible and effective in a clinical setting [90]. The absence of input from radiologists may result in AI tools being designed without factoring in real-world clinical workflows, thereby restricting their usefulness, or posing potential risks to patient safety. Conversely, without the technical prowess of AI experts, radiologists would encounter challenges in harnessing the immense potential of AI for analyzing medical images.

Combining different areas of expertise is essential for successful research. When radiologists and AI developers work together, they can create AI tools that better meet the needs of patients and healthcare professionals [91]. This collaboration also fosters a shared sense of accountability, leading to improved adoption and optimization of AI tools in clinical environments. One effective approach for promoting collaboration is the Academia-Industry Collaboration Plan, which establishes guidelines for partnerships between academic institutions and industry. This is particularly important because universities contribute skilled professionals and new concepts, while the industry offers the financial resources needed for research and innovation.

Nonetheless, the collaboration between academia and industry encounters obstacles. A major challenge that can impede innovation and hinder the implementation of AI tools in radiology is the existing divide between these two sectors. While academia primarily focuses on theory and exploration, industry prioritizes practical applications and market viability [92]. However, there are various approaches that can be utilized to overcome this hurdle. To begin with, collaborative projects involving both academic and industry partners can facilitate the exchange of ideas and resources, resulting in inventive and reliable AI solutions. The utilization of shared datasets enables a wider range of research, thereby improving the applicability of AI tools. Lastly, opensource software provides a platform for cooperation, thereby promoting transparency and reproducibility, which are essential for scientific advancement.

Training in healthcare

The integration of AI into medical practice requires the acquisition of fresh expertise. Given that AI algorithms can handle vast amounts of data surpassing human capabilities, the importance of memorizing extensive medical knowledge or honing procedural abilities through repetitive practice might decrease [93]. This transition necessitates clinicians to gain additional proficiencies, including data science, statistics, and AI ethics, which will play a vital role in interacting with AI technologies securely and efficiently. These proficiencies will enable clinicians to proficiently input data, analyze algorithmic results, and effectively communicate AI-generated treatment strategies to patients.

Given the changing landscape, radiologists may have to go beyond just interpreting images and take on a wider range of responsibilities related to AI technologies. Their focus may shift from solely interpreting images to tasks like developing, validating, and monitoring AI models [94, 95]. An important aspect of AI in radiology involves the labeling of images, a process that requires the expertise of radiologists but is labor-intensive and expensive. As AI technologies progress, radiologists will increasingly work with decision support systems that can provide suggestions for diagnoses, alert test results, and even automate clinical documentation.

The professional growth of radiologists may also align with significant advancements in radiomics and pathomics, enabling the integration of diagnostic services and personalized medicine. This is particularly crucial in settings with limited resources and infrastructure, where the adoption of AI might present challenges [2, 96].

RANZCR has taken a forward-thinking stance towards these obstacles and possibilities. In their updated curriculum for 2023, they have integrated subjects related to AI, showcasing the increasing acknowledgement of AI’s significance in radiology and the necessity for radiologists to develop new skills [97-99]. This pivotal integration of AI guarantees that RANZCR stays ahead in terms of technology and innovative solutions, affirming its dedication to providing radiologists with the latest tools and expertise in the ever-changing realm of medical imaging.

This review concludes by summarizing the main insights, transformative possibilities, and future direction of the complex relationship between AI and medical imaging. With AI playing a crucial role in modern radiology, it offers numerous benefits such as enhanced accuracy in diagnosis, more efficient workflows, and personalized care for patients. These advancements, which include tools for segmenting and categorizing images, computer-assisted diagnosis, and innovative diagnostic and prognostic tools driven by radiomics and predictive analytics, indicate a promising potential for improving patient outcomes. However, there are still challenges to be addressed regarding the privacy and security of data, as well as the opaque nature of AI models. Despite these obstacles, the future looks promising with the development of new algorithms and architectures that expand the scope of medical image analysis. It is essential to foster interdisciplinary research and bridge the gap between academia and industry by promoting collaboration between radiologists and AI developers. This collaboration should also extend to preparing healthcare professionals for an AI-driven landscape and redefining the role of radiologists in the era of AI. Embracing the potential of AI in shaping radiology requires not only a commitment to innovation and the development of advanced algorithms but also nurturing collaborations among radiologists, AI developers, patients, and policymakers. These joint efforts should aim to meet the clinical needs, translate research into practical applications, and ensure the ethical deployment of AI, always prioritizing the safety, privacy, and dignity of patients. In this context, the upcoming era of AI in radiology, though challenging, reveals its immense potential in transforming healthcare.

None.

None.

1. Tang X (2019) The role of artificial intelligence in medical imaging research. BJR Open 2: 20190031. https://doi.org/10.1259/bjro.20190031
2. Najjar R (2023) Redefining Radiology: a review of artificial intelligence integration in medical imaging. Diagnostics 13: 2760. https://doi.org/10.3390/diagnostics13172760
3. Kelly BS, Judge C, Bollard SM, Clifford SM, Healy GM, et al. (2022) Radiology artificial intelligence: a systematic review and evaluation of methods (RAISE). Eur Radiol 32: 7998-8007. https://doi.org/10.1007/s00330-022-08784-6
4. European Society of Radiology (ESR); European Federation of Radiographer Societies (EFRS). Patient safety in medical imaging: A joint paper of the European Society of Radiology (ESR) and the European Federation of Radiographer Societies (EFRS). Insights Imaging 10: 45. https://doi.org/10.1186/s13244-019-0721-y.
5. Dreyer KJ, Geis JR (2017) When machines think: radiology’s next frontier. Radiology 285: 713-718. https://doi.org/10.1148/radiol.2017171183
6. Bushberg JT, Boone JM (2011) The essential physics of medical imaging. Lippincott Williams & Wilkins
7. Ehrlich K, Parker HE, McNicholl DK, Reid P, Reynolds M, et al. (2020) Demonstrating the use of optical fibres in biomedical sensing: a collaborative approach for engagement and education. Sensors 20: 402. https://doi.org/10.3390/s20020402
8. Edler I (1954) The use of ultrasonic reflectoscope for the continuous recording of the movements of heart walls. Kungl Fysiogr Sallsk i Lund Forhandl 24: 1-9. https:// doi.org/10.1570009751004416512
9. Lauterbur, P.C (1973) Image formation by induced local interactions: examples employing nuclear magnetic resonance. Nature 242: 190-191. https://doi. org/10.1038/242190a0 ‘ 1
10. Mansfield P, Grannell PK (1973) J Phys C Solid State Phys 6: L422. https://doi. org/10.1088/0022-3719/6/22/007
11. Huang HK (2019) Pacs-based multimedia imaging informatics: basic principles and applications. John Wiley & Sons.
12. Cherry SR, Jones T, Karp JS, Qi J, Moses WW, et al. (2018) Total-body PET: maximizing sensitivity to create new opportunities for clinical research and patient care. J Nucl Med 59: 3-https://doi.org/10.2967/jnumed.116.184028
13. Hutton BF, Buvat I, Beekman FJ (2011) Review and current status of SPECT scatter correction. Phys Med Biol 56: R85. https://doi.org/10.1088/0031-9155/56/14/R01
14. Vannan MA, Pedrizzetti G, Li P, Gurudevan S, Houle H, et al. (2005) Effect of cardiac resynchronization therapy on longitudinal and circumferential left ventricular mechanics by velocity vector imaging: Description and initial clinical application of a novel method using high?frame rate B?mode echocardiographic images. Echocardiography: A J Cardiovas Ultra Allied Tech 22: 826-830. https://doi.org/10.1111/ j.1540-8175.2005.00172.x
15. Delbeke D, Coleman RE, Guiberteau MJ, Brown ML, Royal HD, et al. (2006) Procedure guideline for SPECT/CT imaging 1.0. J Nucl Med 4: 1227-1234.
16.  Lorenz, J (2016) Management of Malignant Biliary Obstruction Semin Interv Radiol 33: 259-267. https://doi.org/10.1055/s-0036-1592330
17. Uppot RN, Laguna B, McCarthy CJ, De Novi G, Phelps A, et al. (2019) Implementing virtual and augmented reality tools for radiology education and training, communication, and clinical care. Radiology 291: 570-580. https://doi.org/10.1148/ radiol.2019182210
18. von Ende E, Ryan S, Crain MA, Makary MS (2023) Artificial intelligence, augmented reality, and virtual reality advances and applications in interventional radiology. Diagnostics 13: 892 https://doi.org/10.3390/diagnostics13050892
19. Mun SK, Wong KH, Lo SC, Li Y, Bayarsaikhan S (2021) Artificial intelligence for the future radiology diagnostic service. Front Mol Biosci 7: 614258. https://doi. org/10.3389/fmolb.2020.614258
20. Dikici E, Bigelow M, Prevedello LM, White RD, Erdal BS (2020) Integrating AI into radiology workflow: Levels of research, production, and feedback maturity. J Med Imaging 7: 016502.
21. Ahmed N, Abbasi MS, Zuberi F, Qamar W, Halim MS, et al. (2021) Artificial intelligence techniques: analysis, application, and outcome in dentistry-a systematic review. BioMed Res Int 2021: 9751564. https://doi.org/10.1155/2021/9751564
22. Buchanan BG, Shortliffe EH (1984) Rule based expert systems: the mycin experiments of the stanford heuristic programming project (the Addison-Wesley series in artificial intelligence). Addison-Wesley Longman Publishing Co. Inc.
23. Shortliffe E (2012) Computer-based medical consultations: MYCIN. Elsevier.
24. Quinlan JR (1986) Induction of decision trees. Mach Learn 81-106. https://doi. org/10.1007/BF00116251
25. Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 273-297. https:// doi.org/10.1007/BF00994018
26. Rumelhart DE, Hinton GE, Williams RJ (1986) Learning representations by back-propagating errors. Nature 323: 533-536. https://doi.org/10.1038/323533a0
27. LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521: 436-444.
28. Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classification with deep convolutional neural networks. Commun ACM 60: 84-90. https://doi. org/10.1145/3065386
29. Russakovsky O, Deng J, Su H, Krause J, Satheesh S, et al. (2015) Imagenet large scale visual recognition challenge. Int J Comput Vis 115: 211-252. https://doi. org/10.1007/s11263-015-0816-y
30. Thrall JH, Li X, Li Q, Cruz C, Do S, et al. (2018) Artificial intelligence and machine learning in radiology: opportunities, challenges, pitfalls, and criteria for success. J Am Coll Radiol 15: 504-508. https://doi.org/10.1016/j.jacr.2017.12.026
31. Rimmer A (2017) Radiologist shortage leaves patient care at risk, warns royal college. BMJ: British Med J 359. https://doi.org/10.1136/bmj.j4683
32. Bluemke DA, Moy L, Bredella MA, Ertl-Wagner BB, Fowler KJ, et al. (2019) Assessing radiology research on artificial intelligence: a brief guide for authors, reviewers, and readers—from the Radiology Editorial Board. Radiology 294: 487- 489. https://doi.org/10.1148/radiol.2019192515
33. Kahn CE (2019) Artificial intelligence, real radiology. Radiol Artif Intell 1: e184001. https://doi.org/10.1148/ryai.2019184001
34. dos Santos DP, Dietzel M, Baessler B (2021) A decade of radiomics research: are images really data or just patterns in the noise? Eur Radiol 31: 1-4. https://doi. org/10.1007/s00330-020-07108-w
35. Muehlematter UJ, Daniore P, Vokinger KN (2021) Approval of artificial intelligence and machine learning-based medical devices in the USA and Europe (2015-20): a comparative analysis. Lancet Digit Heal 3: e195-e203. https://doi.org/10.1016/ s2589-7500(20)30292-2
36.  Keane PA, Topol EJ (2018) With an eye to AI and autonomous diagnosis. NPJ Digit Med 1: 40. https://doi.org/10.1038/s41746-018-0048-y
37. Wang X, Liang G, Zhang Y (2020) Inconsistent performance of deep learning models on mammogram classification. J Am Coll Radiol 17: 796-803. https://doi. org/10.1016/j.jacr.2020.01.006
38. Jacobson FL, Krupinski EA (2021) Clinical validation is the key to adopting AI in clinical practice. Radiol Artif Intell 3: e210104. https://doi.org/10.1148/ ryai.2021210104
39. Mongan J, Kalpathy-Cramer J, Flanders A, Linguraru MG (2021) RSNA-MICCAI panel discussion: machine learning for radiology from challenges to clinical applications. Radiol Artif Intell 3: e210118. https://doi.org/10.1148/ryai.2021210118
40. Kelly B, Judge C, Bollard SM, Clifford SM, Healy GM, et al. (2020) Radiology artificial intelligence, a systematic evaluation of methods (RAISE): a systematic review protocol. Insights Imaging 11: 1-6. https://doi.org/10.1186/s13244-020-00929-9
41. 41. Ronneberger O, Fischer P, Brox T (2015) U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, Munich, Germany.
42. Arbabshirani MR, Fornwalt BK, Mongelluzzo GJ, Suever JD, Geise BD, et al (2018) Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration. Npj Digit Med 1: 9. https://doi.org/10.1038/s41746-017-0015-z
43. Lebovitz S, Levina N, Lifshitz-Assaf H (2021) Is AI ground truth really true? The dangers of training and evaluating AI tools based on experts’ know-what. Mis Quart 45: 1501-1526. https://doi.org/10.25300/misq/2021/16564
44. Luyckx E, Bosmans JM, Broeckx BJ, Ceyssens S, Parizel PM, et al. (2019) Radiologists as co-authors in case reports containing radiological images: does their presence influence quality? J Am Coll Radiol 16: 526-527. https://doi.org/10.1016/j. jacr.2018.07.035
45. Kelleher JD, Mac Namee B, D’arcy A (2020) Fundamentals of machine learning for predictive data analytics: algorithms, worked examples, and case studies. MIT Press.
46. Blackmore CC (2001) The challenge of clinical radiology research. AJR Am J Roentgenol 176: 327-331. https://doi.org/10.2214/ajr.176.2.1760327
47. Rothstein HR, Sutton AJ, Borenstein M (2005) Publication bias in meta?analysis: Prevention, adjustments. John Wiley & Sons. https://doi.org/10.1002/0470870168. ch10
48. Hussain Z, Gimenez F, Yi D, Rubin D (2017) Differential data augmentation techniques for medical imaging classification tasks. AMIA Annu Symp Proc 2017: 979- 984.
49. Ranschaert ER, Morozov S, Algra PR (2019) Artificial intelligence in medical imaging: opportunities, applications and risks. Springer.
50. Selbst A, Powles J (2017) Meaningful information and the right to explanation. Int Data Privacy Law 7: 233-242.
51. Gunning D, Aha D (2019) DARPA’s explainable artificial intelligence (XAI) program. AI mag. 40: 44-58. https://doi.org/10.1609/aimag.v40i2.2850
52. Keane MT, Smyth B (2020) Good counterfactuals and where to find them: a casebased technique for generating counterfactuals for explainable AI (XAI). In International Conference on Case-Based Reasoning, Salamanca, Spain.
53. Tang A, Tam R, Cadrin-Chênevert A, Guest W, Chong J, et al. 2018. Canadian Association of Radiologists white paper on artificial intelligence in radiology. Can Assoc Radiol J. 69: 120-135. https://doi.org/10.1016/j.carj.2018.02.002
54. Liu X, Rivera SC, Moher D, Calvert MJ, Denniston AK, et al. (2020) Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Lancet Digit Health 2: 537-548. https://doi. org/10.1136/bmj.m3164
55. Mongan J, Moy L, Kahn CE (2020) Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers. Radiol Artif Intell. 2: 200029. https://doi.org/10.1148/ryai.2020200029
56. Korkinof D, Harvey H, Heindl A, Karpati E, Williams G, et al. (2020) Perceived realism of high-resolution generative adversarial network–derived synthetic mammograms. Radiol Artif Intell 3: e190181. https://doi.org/10.1148/ryai.2020190181
57.  Banja J (2020) AI hype and radiology: a plea for realism and accuracy. Radiol Artif Intell 2: e190223. https://doi.org/10.1148/ryai.2020190223
58. Tadavarthi Y, Makeeva V, Wagstaff W, Zhan H, Podlasek A, et al. (2022) Overview of noninterpretive artificial intelligence models for safety, quality, workflow, and education applications in radiology practice. Radiol Artif Intell 4: 210114. https:// doi.org/10.1148/ryai.210114
59. Bizzo BC, Almeida RR, Alkasab TK (2021) Artificial intelligence enabling radiology reporting. Radiol Clin 59: 1045-1052. https://doi.org/10.1016/j.rcl.2021.07.004
60. European Society of Radiology (2019). What the radiologist should know about artificial intelligence–an ESR white paper. Insights imaging 10: 44. https://doi. org/10.1186/s13244-019-0738-2
61. Kumar N, Verma R, Sharma S, Bhargava S, Vahadane A, et al. (2017) A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Trans Med Imaging. 36: 1550-1560. https://doi.org/10.1109/TMI.2017.2677499
62. Komura D, Ishikawa S (2018) Machine learning methods for histopathological image analysis. Comput Struct Biotechnol J 16: 34-42. https://doi.org/10.1016/j. csbj.2018.01.001
63. Attia ZI, Noseworthy PA, Lopez-Jimenez F, Asirvatham SJ, Deshmukh AJ, et al. (2019) An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. Lancet 394: 861-867. https://doi.org/10.1016/S0140-6736(19)31721-0
64. Narula S, Shameer K, Salem Omar AM, Dudley JT, Sengupta PP (2016) Machine- learning algorithms to automate morphological and functional assessments in 2D echocardiography. J Am Coll Cardiol 68: 2287-2295. https://doi.org/10.1016/j. jacc.2016.08.062
65. Zou J, Huss M, Abid A, Mohammadi P, Torkamani A, et al. (2019) A primer on deep learning in genomics. Nat Genet 51: 12-18. https://doi.org/10.1038/s41588- 018-0295-5
66. Can Artificial Intelligence Help See Cancer in New, and Better, Ways? [https:// www.cancer.gov/news-events/cancer-currents-blog/2022/artificial-intelligence-cancer- imaging] [Accessed January 17, 2024].
67. Davenport T, Kalakota R (2019) The potential for artificial intelligence in healthcare. Future Healthc J 6: 94. https://doi.org/10.7861/futurehosp.6-2-94
68. Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, et al. (2018) Scalable and accurate deep learning with electronic health records. NPJ Digit Med 1: 18. https://doi. org/10.1038/s41746-018-0029-1
69.  Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, et al. (2019) A guide to deep learning in healthcare. Nature Med 25: 24-29. https://doi.org/10.1038/ s41591-018-0316-z
70. Qiang B, Chen R, Zhou M, Pang Y, Zhai Y, et al. (2020) Convolutional neural networks- based object detection algorithm by jointing semantic segmentation for images. Sensors 20: 5080. https://doi.org/10.3390/s20185080
71.  Ardila D, Kiraly AP, Bharadwaj S, Choi B, Reicher JJ, et al. (2019) End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nature med. 25: 954-961. https://doi.org/10.1038/s41591-019-0447-x
72. Rodriguez-Ruiz A, Lång K, Gubern-Merida A, Broeders M, Gennaro G, et al. (2000) Stand-alone artificial intelligence for breast cancer detection in mammography: comparison with 101 radiologists. J Natl Cancer Inst 111: 916-922. https://doi. org/10.1093/jnci/djy222
73. Mayo RC, Kent D, Sen LC, Kapoor M, Leung JW, et al. (2019) Reduction of false-positive markings on mammograms: a retrospective comparison study using an artificial intelligence-based CAD. J Digit Imaging 32: 618-624. https://doi. org/10.1007/s10278-018-0168-6
74. AI Converts Low-dose CT Images to High-quality Scans. [https://physicsworld. com/a/ai-converts-low-dose-ct-images-to-high-quality-scans/] [Accessed January 17, 2024].
75. Hickman SE, Baxter GC, Gilbert FJ (2021) Adoption of artificial intelligence in breast imaging: evaluation, ethical constraints and limitations. Br J Cancer 125: 15- 22. https://doi.org/10.1038/s41416-021-01333-w
76. Pesapane F, Rotili A, Agazzi GM, Botta F, Raimondi S, et al. (2021) Recent radiomics advancements in breast cancer: lessons and pitfalls for the next future. Curr Oncol 28: 2351-2372. https://doi.org/10.3390/curroncol28040217
77. Lambin P, Leijenaar RT, Deist TM, Peerlings J, De Jong EE, et al. (2017) Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 14: 749-762. https://doi.org/10.1038/nrclinonc.2017.141
78. Lakhani P, Sundaram B (2017) Deep learning at chest radiography: automated classification of pulmonary tuberculosis by using convolutional neural networks. Radiology 284: 574-582. https://doi.org/10.1148/radiol.2017162326
79. Allen B, Agarwal S, Coombs L, Wald C, Dreyer K (2020) ACR Data Science Institute artificial intelligence survey. J Am Coll Radiol 18: 1153-1159. https://doi. org/10.1016/j.jacr.2021.04.002
80. An Essential Roadmap for AI in Radiology. [https://www.acr.org/Practice-Management- Quality-Informatics/ACR Bulletin/Articles/September-2022/An-Essential- Roadmap-for-AI-in-Radiology] [Accessed January 17, 2024]
81. Yedavalli VS, Tong E, Martin D, Yeom KW, Forkert ND (2021) Artificial intelligence in stroke imaging: Current and future perspectives. clin Imaging 69: 246-254. https://doi.org/10.1016/j.clinimag.2020.09.005
82. Soun JE, Chow DS, Nagamine M, Takhtawala RS, Filippi CG, et al. (2021) Artificial intelligence and acute stroke imaging. Am J Neuroradiol 42: 2-11. https://doi. org/10.3174/ajnr.A6883
83. Artificial Intelligence in MRI. [https://www.umassmed.edu/radiology/radnews/ 2020/10/ai-mri/] [Accessed January 17, 2024]
84. Qiu W, Kuang H, Teleg E, Ospel JM, Sohn SI, et al. (2020) Machine learning for detecting early infarction in acute stroke with non–contrast-enhanced CT. Radiology 294: 638-644. https://doi.org/10.1148/radiol.2020191193
85. Bivard A, Churilov L, Parsons M (2020) Artificial intelligence for decision support in acute stroke—current roles and potential. Nat Rev Neurol 16: 575-585. https:// doi.org/10.1038/s41582-020-0390-y
86. London AJ (2019) Artificial intelligence and black?box medical decisions: accuracy versus explainability. Hastings Cent Rep 49: 15-21.
87. Holzinger A, Langs G, Denk H, Zatloukal K, Müller H (2019) Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov. 9: e1312. https://doi.org/10.1002/widm.1312
88. Barragán-Montero A, Javaid U, Valdés G, Nguyen D, Desbordes P, et al. (2021) Artificial intelligence and machine learning for medical imaging: A technology review. Phys Med 83: 242-256. https://doi.org/10.1016/j.ejmp.2021.04.016
89. Wang G (2016) A perspective on deep imaging. IEEE Access 4: 8914-8924. https:// doi.org/10.1016/j.ejmp.2021.04.016
90. Khosla P, Teterwak P, Wang C, Sarna A, Tian Y, et al. (2020) Supervised contrastive learning. arXiv:2004.11362. https://doi.org/10.48550/arXiv.2004.11362
91. Topol EJ (2019) High-performance medicine: the convergence of human and artificial intelligence. Nature med 25: 44-56. https://doi.org/10.1038/s41591-018-0300-7
92. Murdoch B (2021) Privacy and artificial intelligence: challenges for protecting health information in a new era. BMC Med Ethics 22: 1-5. https://doi.org/10.1186/ s12910-021-00687-3
93. Pianykh OS, Langs G, Dewey M, Enzmann DR, Herold CJ, et al. (2020) Continuous learning AI in radiology: implementation principles and early applications. Radiology 297: 6-14. https://doi.org/10.1148/radiol.2020200038
94. Ahmed F, Fattani MT, Ali SR, Enam RN (2022) Strengthening the bridge between academic and the industry through the academia-industry collaboration plan design model. Front Psychol 13: 875940. https://doi.org/10.3389/fpsyg.2022.875940
95. Banerjee M, Chiew D, Patel KT, Johns I, Chappell D, et al. (2021) The impact of artificial intelligence on clinical education: perceptions of postgraduate trainee doctors in London (UK) and recommendations for trainers. BMC Med Educ 21: 429. https:// doi.org/10.1186/s12909-021-02870-x
96. Clinical Radiology Learning Outcomes and Handbook. [https://www.ranzcr.com/ trainees/clinical-radiology-training-program/learning-outcomes-and-handbook] [Accessed January 17, 2024].
97. Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks. Adv Neural Inf Process 25. https://doi. org/10.1145/3065386