Supervised by Jingxiu Huang

Abstract

Artificial intelligence (AI) is increasingly being used to analyse medical scans, detect signs of disease and flag suspicious findings. This paper examines how its performance compares with that of doctors and whether strong results from medical imaging studies are enough to justify its use in hospitals. Research shows that AI can perform at a similar or higher level than medical professionals on certain well-defined imaging tasks, particularly when identifying specific abnormalities. However, performance on controlled datasets does not necessarily mean reliable diagnosis in everyday clinical settings. AI may struggle when presented with unfamiliar patients, diseases or imaging conditions. For AI to be safely used in hospitals, its performance must be demonstrated beyond the data used during development and tested in realistic clinical environments. These systems also need to be assessed for their limitations and possible errors. Overall, AI shows significant potential to support medical imaging, but only if its performance is thoroughly evaluated and shown to be safe for clinical use.

Research question: How well can AI detect disease in medical scans compared with doctors, and what must be proven before it can be used safely in real hospitals?

1. Introduction

Medical scan reading is a systematic process that translates visual image data into a diagnosis. Artificial intelligence (AI) usage in medical scanning has grown drastically over the last decade, driven by advancements in deep learning (DL). The diagnostic accuracy of AI systems varies greatly depending on the disease, the complexity of the scan and the data used. AI systems excel at detecting specific patterns on scans, with certain models demonstrating higher sensitivity than general radiologists in detecting certain cases. However, AI models can make visual errors, including misinterpreting rare conditions, which is one reason human doctors maintain an advantage over AI systems in complex cases. On the other hand, human specialists face limitations, including fatigue and cognitive biases that can cause missed findings. Before AI can be used in hospitals, a system must not only perform successfully on the data it was trained on, but also perform accurately on independent datasets from entirely different hospitals using different scanner models to ensure that it does not fail in real-world use and to prioritise patient safety.

2. Why Scan Reading is Difficult

2.1 Challenges within the image

Medical images are highly complex, multidimensional data representations of human anatomy. The process of capturing these scans creates barriers that make interpretation difficult before a human observer can interpret them. Medical scans project three-dimensional anatomy onto two-dimensional displays, or present it as a multi-slice series containing hundreds of images, which can obscure pathology. This includes X-rays, which project 3D anatomy onto a 2D display, while computed tomography (CT) and magnetic resonance imaging (MRI) scans produce multi-slice images.

A primary cause of perceptual failure is an abnormality being hidden by the surrounding normal anatomy. In digital frontal chest radiography, an average of 26.4% of the lung volume can be obscured by overlapping structures such as the heart, mediastinum and diaphragm (Chotas & Ravin, 1994). A small, early-stage pulmonary nodule can easily blend into the shadow of a rib. Additionally, the lesion may fail to have a sufficient contrast-to-noise ratio against the background, which can make it visually indistinguishable from normal tissue during a review.

Diagnostic difficulty is often compounded by the fact that different diseases can present with almost identical radiological features. In neuroimaging, distinguishing between a primary glioma, a solitary cerebral metastasis or a tumefactive demyelinating lesion can be difficult on MRI. All three of these conditions can present as a ring-enhancing mass surrounded by prominent vasogenic oedema (Smirniotopoulos et al., 2007). Because their structural features overlap substantially on imaging, relying on morphology alone complicates the differentiation process and increases the risk of misdiagnosis.

Patient movement during image acquisition causes motion blurring and degrades spatial resolution. In coronary computed tomography angiography (CCTA), even small movements from cardiac or respiratory cycles can cause substantial blurring along the coronary arteries, which can compromise the accuracy of coronary lumen boundary detection (Lesser et al., 2009). Similarly, in magnetic resonance cholangiopancreatography (MRCP), respiratory motion can introduce blurring that obscures small stones in the bile duct and reduces the image’s diagnostic utility (Irie et al., 2001).

However, even when an abnormality is visible within an image, radiologists are not scanners, so their interpretation is influenced by search strategies and cognitive heuristics. The majority of diagnostic errors are perceptual rather than interpretive. In a seminal review of diagnostic errors in radiology, Donald and Barnard (2012) analysed 558 diagnostic errors from their sample and noted that 80% were perceptual, meaning that the radiologist failed to identify the abnormality, while 20% were interpretive, meaning that the abnormality was identified but its significance was misjudged, such as dismissing a malignant nodule as a benign scar. These findings suggest that, in this sample, the primary issue in medical image interpretation occurred during the initial detection stage.

2.2 Human interpretation and working conditions

A cognitive bias that compromises detection is known as the ‘satisfaction of the search’ phenomenon (SOS). This phenomenon describes how, once an abnormality is identified, the cognitive drive to search the rest of the scan may drop sharply. For example, if a radiologist detects a prominent acute rib fracture on a trauma patient’s chest CT, they may terminate the systematic review and miss a subtle concurrent pneumothorax. This is because the initial discovery satisfies the diagnostic query. The SOS effect was investigated, and it was found that, during the detection of 23 plain film abnormalities in a dataset of 43 patients, observers missed more plain film abnormalities on contrast studies but also made fewer false positives (Franken et al., 1994). This indicates that missed secondary findings may increase due to a shift in decision-making. On the other hand, AI systems apply a uniform analysis to every image. This approach could flag secondary lesions, which may mitigate the impact of the SOS effect.

2.3 Why these difficulties matter

The combined challenges of structural image complexity, human perceptual blind spots and challenging environments can compromise patient outcomes through inaccurate diagnosis. Because perceptual errors were common in the sample reported by Donald and Barnard (2012), healthcare faces pressure to overcome the limitations of human vision. These persistent challenges have caused researchers to develop AI detection models, which are engineered to evaluate images systematically and may therefore help address some limitations caused by human cognitive biases.

3. How Scan-Reading AI Turns an Image Into an Output

3.1 From a scan to digital information

When you look at a medical scan on a computer screen, it is a highly structured grid of digital information. To analyse a medical image, an AI system requires the scan to be decomposed into fundamental units processed by the computer as numerical matrices. This depends on whether the scan is a flat, two-dimensional image or a three-dimensional volume.

For a two-dimensional medical scan, such as an X-ray, mammogram or digital pathology image, the pixel is the crucial part. A pixel is a picture element and is the smallest piece of data. When an X-ray detector or a digital camera captures an image, it divides the visual image into millions of tiny squares. Each pixel represents a coordinate and holds a specific numerical value. This value corresponds to radiation or light intensity, where a dark pixel indicates a region where more X-rays passed through to reach the detector. A white pixel corresponds to a dense structure, such as bone, that blocked more of the X-ray beam.

Many of the most powerful diagnostic tools, such as computed tomography, magnetic resonance imaging and positron emission tomography (PET), actually capture three-dimensional volumes of the human body and extend the pixel into three dimensions, called a voxel. Each voxel has a defined width, height and depth, and occupies a specific spatial coordinate inside the scanner. In a MRI scan, the voxel value corresponds to the intensity of the radiofrequency signal emitted by relaxing hydrogen protons and depends on the sequence used, reflecting variations in tissue composition, including water and fat. In a 3D medical scan, voxels enable an AI system to analyse volumetric depth. However, processing voxel data puts higher demands on computational resources than processing pixel-based data. A high-resolution chest CT scan normally consists of hundreds of millions of voxels.

3.2 Model types and pattern recognition

An AI system does not immediately recognise the shape of a heart or rib cage; instead, it sees the numbers. For example, if a chest X-ray is 2,000 pixels wide by 2,000 pixels high, the input layer of the AI system processes a numerical matrix which contains 4 million numerical values. Every calculation that the AI performs begins by analysing patterns, gradients and relationships within the numbers. When a medical scan is converted into a grid of pixels, it enters the core of the AI system, which consists of artificial neural networks inspired by biological neural pathways. Specialised network models have been engineered specifically to process complicated visual data. Historically, convolutional neural networks (CNNs) have been used as a primary system of computer vision in healthcare. These models automate feature extraction directly from pixels, which advances the capabilities of computer-aided diagnosis.

CNNs use convolution, which is an operation where a small digital filter slides across the pixels of an image. As it moves, it multiplies its own values by the image values and adds them to create a filtered image called a feature map. They process images locally, so they are efficient at recognising spatial structures. CNNs have successfully been used for diagnostic purposes in hospitals, with strong performance on large patient datasets. In a cohort of 20,530 patients, the ‘PANDA’ system demonstrated a sensitivity of 92.9% and a specificity of 99.9% for pancreatic lesion detection (Cao et al., 2023). The model failed to identify about 7.1% of positive cases at that threshold.

Although CNNs have been a dominant medical AI system, a newer model type known as the vision transformer (ViT) has emerged, inspired by transformer models in natural language processing. A vision transformer processes a scan by cutting a 2D image into square segments, converting them into vectors of numbers and lining them up in a sequence. Each transformer encoder block includes a self-attention mechanism alongside a feed-forward network and layer normalisation. Self-attention allows the model to learn relationships between patches from different parts of the image (Hosny & Mohammed, 2025).

3.3 Training, validation and testing

Once a medical scan has been converted into pixel or voxel numerical inputs, an AI system is trained using training data. During training, the model learns the mapping between image patterns and clinical labels by repeatedly updating its internal parameters to reduce prediction errors. During this stage, performance may look high because the model is learning directly from those specific examples. The model is evaluated on a separate validation set to help adjust how the model is trained, and this can be done by selecting hyperparameters such as the learning rate and number of layers.

Additionally, the best-performing model is chosen to reduce the risk of the AI overfitting to the training set. In medical imaging, this is important because AI systems could achieve excellent results on the data they were shown but generalise poorly if tuning is done incorrectly or if test data is ‘leaked’ into training. The test set is then used once in order to check the model’s performance on cases that were not seen during training or validation. This provides an initial estimation of how well the AI would work in real-world clinical environments, although further external and prospective testing is still needed.

3.4 Outputs and clinical workflow

The format of an output depends on the task the model was built to do, and that task then determines how the result is used inside the clinical workflow. The model may output a single predicted class, such as benign or malignant. Instead of a category, the model may predict a continuous value, such as a severity score or risk score used for monitoring. For tasks such as detecting suspicious findings, the output may include marked locations known as candidate positions to show where a lesion is likely to be located.

For outlining anatomy precisely, such as tumours and organs, the model output is a mask showing which pixels belong to the target structure. Neural networks usually produce a score, often between 0 and 1, which may be interpreted as a probability and used by clinicians to support a decision about whether to treat or flag the case as positive.

A continuous model score can be converted into a decision using a threshold. If the output score is greater than or equal to the threshold, it is classified as positive, but if it is below the threshold, it is classified as negative. The threshold is often selected using validation data to match clinical priorities, such as controlling false positives.

Overall, the model finally outputs a score, sometimes interpreted as a probability, for a diagnosis, the location of a suspected finding or an outlined area of anatomy. The system then converts the score using a threshold into an action. Doctors use these outputs to help their decisions, such as prioritising cases, confirming disease presence or checking locations before making a final diagnosis.

4. How AI Compares with Doctors on Specific Tasks

4.1 What makes a fair comparison

To make a fair comparison between AI and doctors, researchers have to test both on the same complex diagnostic task using the same patient images under the same study conditions. In a landmark study on retinal disease diagnosis, researchers developed an AI model using 14,884 anonymised 3D optical coherence tomography (OCT) scans from 7,621 patients at Moorfields Eye Hospital and compared its referral decisions with experts on an independent test set of 997 patients (De Fauw et al., 2018). The AI reached or exceeded expert performance on referral decisions covering more than 50 retinal diseases. This finding is important because it shows that AI can match top human specialists when given identical complex imaging data. However, this comparison is limited to image-based triage and does not account for a doctor’s ability to communicate with patients in person.

Moreover, in a study published by Stanford University researchers, an AI model called CheXNet was evaluated against four practising Stanford radiologists for pneumonia detection. The full dataset contained 112,120 frontal chest X-ray images from 30,805 unique patients treated at the National Institutes of Health Clinical Centre, while the direct comparison with radiologists used a separate test set of 420 images. When identifying pneumonia, the AI algorithm scored an F1 metric of 0.435 compared with the specialists’ average of 0.387, showing higher performance on this study task (Rajpurkar et al., 2017). This finding is important because it shows that AI can match trained doctors in spotting a potentially life-threatening condition on chest X-rays. However, the algorithm was tested only on static 2D images, meaning it could not access and analyse patient medical histories or perform follow-up clinical checks.

4.2 Evidence across medical imaging tasks

To evaluate how artificial intelligence compares with human doctors across multiple medical imaging tasks, researchers must compare both on identical datasets across different scan modalities rather than in one specialised area. In a global systematic review and meta-analysis led by the University of Birmingham, researchers analysed diagnostic data comparing deep learning algorithms with healthcare professionals across various imaging types, including chest X-rays, brain MRIs and CT scans. Across the 14 studies that directly compared both groups on the same patient images, the AI models achieved a combined sensitivity of 87.0% compared with 86.4% for human doctors, and a combined specificity of 92.5% compared with 90.5% for doctors (Liu et al., 2019). This finding matters for the research question because it shows that AI performance is not an isolated result in one department, but a consistent trend across several medical imaging fields. However, a fair caution is that most of these studies used retrospective image datasets, meaning they do not show how AI performs alongside doctors in live, chaotic clinical environments.

When we compare the above studies, clear patterns about how AI compares with human doctors can be found. In each case, AI met or exceeded human performance in identifying specific structural abnormalities, showed consistency in analysing pixels in high-resolution digital images and had high diagnostic accuracy on the tasks studied.

4.3 What the results can and cannot show

Nevertheless, AI is not entirely superior to human doctors in actual healthcare, but rather a tool to increase efficiency. Firstly, artificial intelligence lacks emotions and human communication skills. It cannot communicate bad news with empathy, answer patient questions with patience or do its absolute best to help patients with compassion. Moreover, even though AI excels in image analysis, it is only one factor in determining a patient’s problem. Human doctors can evaluate a patient’s physical symptoms, review medical histories, notice subtle behavioural cues and adjust decisions dynamically.

Therefore, AI can contribute to the world of medicine, yet it cannot replace the full, emotional and multifaceted role of a human doctor.

The results from studies comparing AI with doctors provide evidence that artificial intelligence can perform at a high level on specific medical imaging tasks. The Moorfields study demonstrated that an AI system could achieve high performance when making referral decisions from optical coherence tomography scans, while the CheXNet study showed that an AI model could achieve performance comparable to radiologists when detecting pneumonia on chest X-rays. These findings suggest that AI has the potential to assist doctors, particularly when the task is clearly defined and the model is given the same images as the human experts.

However, these results do not prove that AI is better than doctors at diagnosis overall. Most comparison studies evaluate a specific task using a particular dataset instead of testing the complete clinical decision-making process. A radiologist can consider a patient’s symptoms, medical history and other information when interpreting the image, while an image-based AI model may receive only the scan itself. Therefore, higher performance on an imaging benchmark does not mean that AI can replace a doctor.

The results also cannot guarantee that an AI system will remain consistent and provide the same performance in every hospital. Models may perform differently when they encounter new scanners, imaging protocols, patient populations or disease patterns that were not well represented in their original training data. This means that strong internal test results demonstrate what an AI system can achieve under the conditions in which it was evaluated, but they do not by themselves show that the system is reliable or safe for widespread use. External validation and testing in realistic environments are necessary before conclusions can be drawn about the system’s effectiveness in real hospitals.

5. Limits and Performance in Real Hospitals

5.1 Generalisation to new hospitals and patients

A systematic review found that AI generally performed better in its development setting than in external settings (Suleman et al., 2025). Out of the 342 records identified, only six studies met the requirements; the others were excluded because they were duplicates, did not use external validation, were outside the radiology domain or lacked sufficient metrics. In this review, external validation showed lower performance than internal validation. This means that AI worked well with familiar data, but performance decreased when it was tested somewhere new: the median AUC decreased by about 0.03, while specificity decreased by as much as 24 percentage points in some studies (Suleman et al., 2025). This reduced performance may occur because hospitals have different scanners, patient age groups, imaging protocols or disease patterns. This can create a dataset shift for the AI. Reduced specificity could lead to more false-positive results and unnecessary follow-ups, increasing the workload for healthcare workers, while reduced sensitivity could increase false-negative results and potentially cause something important to be missed. This review included only six studies, and none prospectively tested AI after deployment in a real hospital setting, so it does not mean that AI will fail when moved to another hospital. Instead, the findings identify a risk of reduced performance when AI is transferred to new hospitals rather than stating that every AI model will fail outside its development setting.

5.2 Bias, labels and shortcut learning

It is possible for AI to produce biased results when there are limitations in its training data or labels because it learns patterns that do not generalise fairly across all patients. If an AI model is trained on a dataset that consists mostly of images from a particular racial, age or ethnic group, it would make sense for there to be a reduction in the AI’s accuracy when it is applied to a different group. By being given less representative data, the AI can unintentionally produce biased results (Kocak et al., 2025). Ultimately, removing humans from one part of the decision-making process in the healthcare system does not necessarily eliminate human bias, since the whole process starts with human decisions. Some of that data may be biased; the AI may therefore learn from those patterns and then make predictions based on them. The question becomes: is AI biased, or is it the information used by AI and the assumptions or biases already embedded in that information?

AI can learn from patterns in medical data that appear useful for making a prediction but are not actually related to the disease. Across 13 datasets and 207,487 patients, researchers found that model performance was often overestimated by up to 20% because of shortcut learning connected to hidden data acquisition biases (Ong Ly et al., 2024). This research explains the differences in how medical images are acquired. Different scanners or imaging settings can create signals that an AI can detect even though those signals are not part of the disease. Those differences can later become shortcuts for the AI. For example, an AI could use differences in X-ray contrast as a shortcut rather than focusing on the patient’s anatomy. A model achieved an AUROC of 0.85 when it was evaluated on the MIMIC-CXR dataset, but its performance dropped to 0.73 when tested on the external CheXpert dataset (Ong Ly et al., 2024). This shows that the model’s strong performance on its original data did not necessarily mean that it had learnt clinical features related to the disease; instead, it may have relied on shortcuts that were specific to its environment. Overall, AI can be correct for the wrong reasons.

5.3 Clinical consequences of errors

In a clinical setting, an AI error can greatly influence the care a patient receives. Depending on how an AI reads the scan, a false positive and a false negative can have different consequences for a patient’s diagnosis. A false negative occurs when AI does not flag a disease that is present. This could cause a delayed diagnosis or an unrecognised disease, which could be dangerous for the patient’s health. A false positive occurs when AI flags a disease that is not present. This could lead to additional investigation, unnecessary testing and pressure on medical resources (Evans & Snead, 2024). These consequences show that AI performance cannot be evaluated solely on its overall accuracy. The risks of reduced generalisation, biased or unrepresentative training data, shortcut learning and clinically significant errors show that strong performance during development does not guarantee safe performance in real-world healthcare. These limitations show why safeguards are needed when using AI in healthcare.

6. Safe Use, Trust and Responsibility

6.1 Evidence before clinical use

Strong results during development alone are not enough to demonstrate that an AI system is safe for clinical use because AI performance can decrease when it is tested in unfamiliar hospitals and patient populations. Before an AI diagnostic tool is introduced into a hospital, its technical performance is validated to determine how accurately it performs its intended task. Even if it passes technical validation, medical professionals will still have to conduct further clinical validation in a realistic clinical setting (Park et al., 2021). They also distinguish between technical performance and clinical utility because an AI system may have high diagnostic accuracy without necessarily improving patient outcomes. Therefore, before AI can be trusted in a real hospital setting, evidence should demonstrate not only that it can accurately detect and flag disease, but also that it generalises to the patient population and clinical setting for which it is intended.

6.2 Human oversight and ongoing monitoring

A recent review of human-in-the-loop AI in healthcare found that combining human expertise with AI can support safer clinical decision-making while maintaining accountability (Olawade et al., 2026). This is extremely important in medical imaging because AI can produce uncertain or incorrect results, meaning doctors need to be able to recognise when an AI recommendation should be questioned rather than automatically accepting it. Even after an AI system has been clinically validated, it should not make medical decisions independently without human oversight because its performance can change as patient populations, equipment, environments and clinical practices change.

6.3 Fairness, regulation and responsibility

Before an AI system is introduced into a hospital, several safeguards should be established to ensure that it is safe and reliable. Since AI performance can decrease when it is tested in unfamiliar hospitals, the system should undergo external validation using data from different hospitals, scanners and patient populations before being used. AI should also be tested across diverse patient groups because limitations in training data can cause biased performance (Tripathi et al., 2023). Since false positives and false negatives can affect patient care, the system’s errors should be evaluated to determine whether its performance is acceptable for its intended use. After implementation, hospitals should continue monitoring the AI to identify changes in performance or unexpected errors. Ultimately, healthcare professionals, hospitals and AI developers should have defined roles in responding to AI errors, while doctors maintain oversight of the AI’s recommendations.

7. Conclusion

To conclude, artificial intelligence is a remarkably useful tool in modern medicine, but it does not replace the broader role of human physicians. AI stands out for its ability to process large volumes of data and identify subtle patterns in medical scans. However, practising medicine requires far more than analysing digital information. A computer program cannot offer genuine empathy, reassure an anxious patient or apply critical thinking during unpredictable clinical emergencies. AI serves best as an advanced assistant by handling complex data tasks so that doctors can focus on providing compassionate, high-quality patient care. Safe use in hospitals therefore requires external validation, human oversight and ongoing monitoring.

Cao, K., Xia, Y., Yao, J. et al. (2023) ‘Large-scale pancreatic cancer detection via non-contrast CT and deep learning’, Nature Medicine, 29, pp. 3033-3043.

Chotas, H.G. and Ravin, C.E. (1994) ‘Chest radiography: estimated lung volume and projected area obscured by the heart, mediastinum, and diaphragm’, Radiology, 193(2), pp. 403-404.

De Fauw, J., Ledsam, J.R., Romera-Paredes, B. et al. (2018) ‘Clinically applicable deep learning for diagnosis and referral in retinal disease’, Nature Medicine, 24, pp. 1342-1350.

Donald, J.J. and Barnard, S.A. (2012) ‘Common patterns in 558 diagnostic radiology errors’, Journal of Medical Imaging and Radiation Oncology, 56(2), pp. 173-178.

Evans, H. and Snead, D. (2024) ‘Understanding the errors made by artificial intelligence algorithms in histopathology in terms of patient impact’, npj Digital Medicine, 7, article 89.

Franken, E.A. Jr, Berbaum, K.S., Lu, C.H. et al. (1994) ‘Satisfaction of search in the detection of plain-film abnormalities in abdominal contrast studies’, Investigative Radiology, 29(4), pp. 403-409.

Hosny, K.M. and Mohammed, M.A. (2025) ‘Explainable AI and vision transformers for detection and classification of brain tumor: a comprehensive survey’, Artificial Intelligence Review, 58, article 259.

Irie, H., Honda, H., Kuroiwa, T. et al. (2001) ‘Pitfalls in MR cholangiopancreatographic interpretation’, RadioGraphics, 21(1), pp. 23-37.

Kocak, B., Ponsiglione, A., Stanzione, A. et al. (2025) ‘Bias in artificial intelligence for medical imaging: fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects’, Diagnostic and Interventional Radiology, 31(2), p. 75.

Lesser, J.R., Flygenring, B., Knickelbine, T., Longe, T. and Schwartz, R.S. (2009) ‘Practical approaches to overcoming artifacts in coronary CT angiography’, Journal of Cardiovascular Computed Tomography, 3(1), pp. 4-15.

Liu, X., Faes, L., Kale, A.U. et al. (2019) ‘A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis’, The Lancet Digital Health, 1(6), e271-e297.

Olawade, D.B., Plabon, S.B., Ojo, A. et al. (2026) ‘Human in the loop artificial intelligence in healthcare: applications, outcomes, and implementation challenges’, International Journal of Medical Informatics, article 106362.

Ong Ly, C., Unnikrishnan, B., Tadic, T. et al. (2024) ‘Shortcut learning in medical AI hinders generalization: method for estimating AI model generalization without external data’, npj Digital Medicine, 7, article 124.

Park, S.H., Choi, J. and Byeon, J.S. (2021) ‘Key principles of clinical validation, device approval, and insurance coverage decisions of artificial intelligence’, Korean Journal of Radiology, 22(3), pp. 442-453.

Rajpurkar, P., Irvin, J., Zhu, K. et al. (2017) ‘CheXNet: radiologist-level pneumonia detection on chest X-rays with deep learning’, arXiv:1711.05225.

Smirniotopoulos, J.G., Murphy, F.M., Rushing, E.J., Rees, J.H. and Schroeder, J.W. (2007) ‘Patterns of contrast enhancement in the brain and meninges’, RadioGraphics, 27(2), pp. 525-551.

Suleman, M.U., Mursaleen, M., Khalil, U. et al. (2025) ‘Assessing the generalizability of artificial intelligence in radiology: a systematic review of performance across different clinical settings’, Annals of Medicine and Surgery, 87(12), p. 8803.

Tripathi, S., Gabriel, K., Dheer, S. et al. (2023) ‘Understanding biases and disparities in radiology AI datasets: a review’, Journal of the American College of Radiology, 20(9), pp. 836-841.