A Literature Review on Deep Learning for Medical Image Diagnosis

Introduction

Medical imaging sits at the centre of modern diagnostic practice, and the volume of images now generated across radiology, pathology and ophthalmology has grown faster than the clinical workforce available to interpret them. Deep learning, a family of representation-learning methods built on multi-layered neural networks, has emerged over the past decade as the dominant computational approach to automating and augmenting this interpretation (LeCun, Bengio and Hinton, 2015). Where earlier computer-aided diagnosis systems depended on hand-engineered features designed by domain experts, deep networks learn hierarchical features directly from pixel data, a shift that has coincided with rapid and well-documented gains in diagnostic task performance (Litjens et al., 2017). The scale of clinical interest is now such that regulators have begun to approve imaging algorithms for routine use, and the number of cleared devices has grown into the hundreds, the majority of them concentrated in radiology (Benjamens, Dhunnoo and Mesko, 2020). This transition from research prototype to regulated product raises questions that a purely technical account cannot answer, and it makes a careful reading of the evidence base more, not less, important.

This review aims to synthesise the existing scholarly literature on the application of deep learning to medical image diagnosis, rather than to report any new experimental findings. Its scope is deliberately bounded. It concentrates on supervised deep learning applied to two-dimensional and volumetric diagnostic imaging in four well-studied clinical domains: radiology, pathology, ophthalmology and dermatology. It considers how performance has been evaluated through benchmark studies, and it examines the recurring obstacles that the literature identifies, namely data availability, generalisation, interpretability, bias and the emerging demands of regulation. Generative modelling, image reconstruction and treatment-planning applications fall outside the present scope. The review proceeds by first establishing the technical background, then thematically synthesising applications, performance evidence and challenges, before offering a critical appraisal of gaps and a short conclusion. Throughout, the guiding question is not merely what deep learning can achieve under laboratory conditions, but under what conditions the reported achievements can be trusted to transfer to clinical practice.

Background: Convolutional Neural Networks and Medical Imaging

The convolutional neural network (CNN) is the architectural backbone of most medical image diagnosis research. CNNs exploit the spatial structure of images through local receptive fields, shared weights and pooling operations, allowing them to learn translation-tolerant features while keeping the number of parameters tractable (Goodfellow, Bengio and Courville, 2016). The contemporary interest in CNNs is often traced to the marked improvement they produced on the ImageNet natural-image classification benchmark, where a deep architecture substantially outperformed prior approaches (Krizhevsky, Sutskever and Hinton, 2012). Subsequent architectural refinements, including very deep networks with small convolutional filters (Simonyan and Zisserman, 2015), residual connections that ease the training of extremely deep models (He et al., 2016), and densely connected blocks that promote feature reuse and gradient flow (Huang et al., 2017), established design patterns that medical imaging research adopted with relatively little modification. The consistency of this borrowing is itself notable: rather than developing bespoke architectures, much medical imaging work has reused general-purpose vision backbones and concentrated its innovation on data curation, training regimes and task framing.

Two features of the medical domain shaped how these architectures were applied. First, labelled medical datasets are typically far smaller than natural-image corpora, which encouraged the widespread use of transfer learning, where networks pre-trained on general images are fine-tuned on clinical data (Shen, Wu and Suk, 2017). Transfer learning is not without cost, and later work has questioned how much a network pre-trained on natural photographs genuinely brings to a domain whose textures and structures differ markedly from everyday scenes; nonetheless it remains the default starting point in most pipelines. Second, many diagnostic tasks require segmentation, the pixel-level delineation of anatomical or pathological regions, rather than whole-image classification. The U-Net architecture, with its symmetric encoder-decoder structure and skip connections, was proposed specifically for biomedical segmentation and has become a de facto standard in the field (Ronneberger, Fischer and Brox, 2015). More recent work has shown that careful, automated configuration of such a network, rather than architectural novelty, is frequently the decisive factor in segmentation performance, a finding that reinforces the broader lesson that engineering discipline often matters more than model exotica (Isensee et al., 2021).

A newer strand of architectural research has begun to reshape the field. The attention mechanism, originally developed for sequence modelling, dispenses with convolution’s fixed locality in favour of learned, content-dependent relationships between all parts of an input (Vaswani et al., 2017). Its adaptation to images, the vision transformer, treats an image as a sequence of patches and has been reported to match or exceed convolutional baselines when sufficient data are available (Dosovitskiy et al., 2021). Whether transformers offer a durable advantage in the comparatively data-poor medical setting remains an open question in the literature, and much recent work explores hybrid designs. Alongside architecture, self-supervised pre-training, in which representations are learned from unlabelled images before any diagnostic labels are introduced, has attracted attention as a means of exploiting the large volumes of unannotated clinical imagery that hospitals accumulate (Azizi et al., 2021). The comprehensive survey by Litjens et al. (2017) documents how the earlier building blocks were assembled across hundreds of studies, and it remains a common reference point for situating this later work.

Thematic Synthesis

The literature can be organised as a chain running from the diagnostic tasks that deep learning has been applied to, through the technical and data-related factors that have enabled progress, to the barriers that limit clinical translation, and finally to the research gap those barriers define. Figure 1 summarises this structure, and the sections that follow examine each element in turn.

Applications across clinical domains

In radiology, deep learning has been applied to modalities ranging from chest radiography to computed tomography and magnetic resonance imaging. A frequently cited example is the detection of pathology on chest X-rays, where large publicly released datasets enabled the training of networks to flag conditions such as pneumonia and other thoracic abnormalities (Rajpurkar et al., 2017). Volumetric applications include the analysis of low-dose computed tomography for lung cancer screening, where end-to-end models were trained to assess whole scans (Ardila et al., 2019). Mammography has attracted particular attention because of established population screening programmes; retrospective work has evaluated deep models against radiologist reads on large screening cohorts (McKinney et al., 2020). What unites these examples is the combination of a high-stakes, high-volume screening task with a modality for which sizeable annotated datasets could be assembled, and it is precisely that combination, rather than any single algorithmic breakthrough, that has driven progress.

In pathology, the digitisation of glass slides into gigapixel whole-slide images created both an opportunity and a computational challenge. A single slide may contain billions of pixels, which forces most pipelines to divide the image into smaller tiles, classify each tile, and then aggregate the results into a slide-level decision. Studies have trained CNNs to classify tumour subtypes and predict molecular features directly from histopathology tiles (Coudray et al., 2018), and to detect the metastatic spread of breast cancer to lymph nodes, a task formalised in the widely used CAMELYON challenge and its associated benchmark comparison against pathologists (Ehteshami Bejnordi et al., 2017). Because pixel-level annotation of enormous slides is costly, weakly supervised approaches that learn from slide-level diagnostic labels alone have been proposed to make clinical-scale training feasible (Campanella et al., 2019). This design choice recurs throughout the pathology literature and illustrates a broader theme: architectural and training decisions in medical imaging are frequently driven as much by the practical economics of annotation as by considerations of raw model capacity. The same economic pressure explains the growing interest in self-supervised pre-training, which promises to convert vast archives of unlabelled slides into useful representations without the annotation bottleneck (Azizi et al., 2021).

Ophthalmology has proved an especially productive testbed, in part because retinal fundus photographs are standardised and abundant. Deep learning systems have been developed to detect diabetic retinopathy from fundus images (Gulshan et al., 2016), with parallel work validating similar systems across multi-ethnic populations (Ting et al., 2017). Extending beyond a single disease, De Fauw et al. (2018) reported a system using optical coherence tomography to triage a range of retinal conditions and referral decisions. Dermatology follows a comparable pattern: Esteva et al. (2017) trained a network to classify skin lesions, including malignant melanoma, from clinical and dermoscopic photographs, framing the task as a direct comparison with board-certified specialists. Across all four domains, the recurring pattern is one of narrowly scoped, well-defined diagnostic tasks paired with comparatively large curated datasets. Table 1 draws these applications together, mapping each clinical domain to its characteristic imaging modality and representative diagnostic tasks.

The table also makes an omission visible. The four domains that dominate the literature share standardised acquisition, high image volumes and, in most cases, a single well-posed classification target. Domains that lack these properties, whether because imaging is more heterogeneous, disease prevalence is lower, or annotation requires rarer expertise, are comparatively under-studied, a point returned to in the critical appraisal.

Performance and benchmark studies

The evidence base for deep learning in medical imaging rests heavily on benchmark and comparison studies, and the literature is careful about how performance is framed. Rather than treating any single reported metric as definitive, several authors emphasise comparison against clinician performance on matched tasks and datasets. Esteva et al. (2017), for instance, evaluated a dermatological image classifier against board-certified specialists, and studies in ophthalmology and pathology adopted analogous designs (Gulshan et al., 2016; Ehteshami Bejnordi et al., 2017). The metrics most commonly reported, the area under the receiver operating characteristic curve together with sensitivity and specificity at a chosen operating point, are informative but incomplete. An area-under-curve figure summarises discrimination across all thresholds and can appear reassuringly high even when the clinically relevant operating point performs poorly, and sensitivity and specificity trade off against one another in ways that depend on disease prevalence and on the costs of false positives and false negatives in a given care pathway. Reviews of the clinical evidence have accordingly cautioned that many such studies are retrospective, single-site and reported using heterogeneous metrics, which complicates cross-study comparison (Topol, 2019). A further subtlety concerns the reference standard against which models are judged. Where the ground truth is itself derived from expert reads rather than from an independent outcome such as biopsy or long-term follow-up, a model that agrees closely with the reference may be rewarded for replicating human error as much as for detecting disease, and the ceiling on measurable performance is set by the reliability of the labels themselves. Reported figures are therefore best read as promising signals rather than as settled measures of clinical accuracy, and headline comparisons with clinicians are most informative when the clinicians in question worked under conditions, and with information, comparable to routine practice.

A more pointed methodological critique concerns the design and reporting quality of diagnostic accuracy studies. A systematic review by Liu et al. (2019) compared deep learning performance against health-care professionals across many imaging tasks and found that comparatively few studies used external validation or prospective designs, and that reporting standards varied considerably. The same review noted that a substantial proportion of published work was excluded from quantitative synthesis precisely because it did not report results in a comparable form, which suggests that the visible performance literature may be both selective and optimistic. A related concern is publication bias: positive results are more likely to be written up and accepted, so the aggregate impression conveyed by the literature may overstate typical performance. This has motivated the development of reporting and trial guidelines tailored to artificial intelligence interventions, including extensions to established trial-reporting frameworks that specify what an AI intervention study must disclose (Liu et al., 2020). Several commentators regard such guidelines as necessary for the field to mature from proof-of-concept demonstrations towards clinically trustworthy evidence (Kelly et al., 2019). Taken together, the literature therefore presents strong performance signals alongside persistent doubts about whether reported figures generalise beyond the narrow and often favourable conditions under which they were obtained.

Challenges: data, generalisation, interpretability, bias and regulation

The first cluster of challenges concerns data. Supervised deep learning is dependent on large, accurately labelled datasets, yet medical annotations require scarce expert time, and privacy, governance and interoperability constraints restrict data sharing (Willemink et al., 2020). Class imbalance, inconsistent labelling protocols and variation in imaging equipment further complicate dataset construction. These conditions have made transfer learning, data augmentation and, more recently, self-supervised and weakly supervised strategies central to the field (Campanella et al., 2019; Shen, Wu and Suk, 2017; Azizi et al., 2021). One response to the tension between data hunger and data governance is federated learning, in which a model is trained across several institutions without the underlying images ever leaving their home sites; proponents argue that this can widen the effective training distribution while respecting privacy, though it introduces its own engineering and governance complexity (Rieke et al., 2020). Data quality, moreover, is not simply a matter of quantity: labels generated by a single institution’s convention may encode systematic idiosyncrasies that a model faithfully, and unhelpfully, reproduces. The provenance of the imaging data matters as well, since a corpus assembled retrospectively from routine care may over-represent the populations, scanners and referral patterns of the contributing institutions, so that a dataset which is large in absolute terms is nonetheless narrow in the diversity that generalisation demands.

The second challenge, generalisation, follows closely. Models frequently perform worse when evaluated on data from institutions, scanners or populations not represented in training, a manifestation of distribution shift. An early and influential demonstration found that a pneumonia-detection model trained at one set of hospitals generalised poorly to another, partly because it had learned to infer hospital-specific factors from image characteristics rather than the pathology itself (Zech et al., 2018). A related and widely discussed illustration is the tendency of networks to exploit spurious correlations, or shortcut features, such as site-specific markers rather than genuine pathology; DeGrave, Janizek and Lee (2021) demonstrated this in the context of COVID-19 detection from chest radiographs, where models latched onto dataset artefacts rather than disease signs. Compounding these concerns, medical models have been shown to be vulnerable to small, deliberately crafted perturbations that alter a prediction while remaining imperceptible to a human reader, raising security as well as reliability questions (Finlayson et al., 2019). Such findings underline why external validation is treated in the literature as a prerequisite for clinical claims rather than as an optional robustness check.

The third challenge is interpretability. Deep networks are frequently characterised as opaque, which is problematic in a domain where clinicians and regulators require justifiable decisions. A substantial body of work has developed post-hoc explanation techniques, including gradient-based class-activation mapping that highlights image regions influencing a prediction (Selvaraju et al., 2017). However, critics argue that saliency-style explanations can be unstable and may offer a false sense of transparency, and that inherently interpretable models or rigorous validation may be preferable to post-hoc rationalisation (Rudin, 2019). The difficulty is not only technical but epistemic: a heat-map indicating where a network looked does not establish why it reached its conclusion, nor whether the highlighted region reflects genuine pathology or an incidental correlate. Interpretability thus remains contested rather than solved, and the debate maps closely onto the generalisation problem, since an explanation that reveals shortcut behaviour is only useful if practitioners act on what it reveals.

The fourth challenge, bias and equity, cuts across the others. Because models inherit the characteristics of their training data, under-representation of particular demographic groups can produce systematically unequal performance across, for example, skin tone, age or geographic origin. Although the most prominent demonstration of algorithmic bias in health concerned a risk-prediction model rather than an imaging system, its lesson, that proxies and skewed data encode inequity, is treated as directly applicable to imaging (Obermeyer et al., 2019). The concern is sharpened by the phenomenon of hidden stratification, in which aggregate performance masks poor performance on clinically important subgroups that are invisible unless results are disaggregated (Oakden-Rayner et al., 2020). It is compounded further by the shortcut-learning findings noted above, since a model that keys on site-specific or demographic artefacts may appear accurate in aggregate while failing for particular subpopulations (DeGrave, Janizek and Lee, 2021). Reviews of clinical translation accordingly position fairness auditing, and the reporting of performance disaggregated by subgroup, alongside overall accuracy as a condition for responsible deployment (Kelly et al., 2019).

A fifth challenge, regulation, has grown in prominence as systems have moved towards deployment. Regulatory bodies have begun to clear imaging algorithms as medical devices, and an analysis of approved products found that the great majority were in radiology and that documentation of the evidence supporting them was often limited (Benjamens, Dhunnoo and Mesko, 2020). Regulation poses questions that the technical literature does not resolve on its own: how a system that continues to learn after deployment should be evaluated, how performance drift over time should be monitored, and what standard of prospective evidence should be required before and after market approval. Reporting guidelines for AI interventions represent one part of the response, translating the methodological critiques discussed above into concrete requirements for what studies and, ultimately, regulatory submissions must disclose (Liu et al., 2020). Regulation therefore functions less as a separate obstacle than as the institutional setting in which the other four challenges must be demonstrably addressed.

Critical Appraisal and Gaps

Read together, the literature exhibits a consistent asymmetry. The volume of studies reporting strong performance on circumscribed diagnostic tasks is large and continues to grow, yet the body of evidence demonstrating robust, prospective, multi-site clinical utility remains comparatively thin. Much of the foundational and most-cited work is retrospective and internally validated, and even landmark studies tend to isolate a single task under favourable data conditions (Litjens et al., 2017; Topol, 2019). This creates a gap between technical feasibility, which is well established, and clinical effectiveness, which is far less so. The pipeline in Figure 2 makes the point concrete: the algorithmic core, the CNN that maps an image to a prediction, is comparatively mature, whereas the surrounding apparatus of validation, oversight and monitoring, on which safe deployment depends, is far less developed in the primary studies.

A second gap concerns methodological consistency. The absence, until relatively recently, of agreed reporting standards means that headline metrics are often not comparable across studies, and negative or null results appear under-represented in the published record (Liu et al., 2019; Kelly et al., 2019). This weakens the field’s capacity to synthesise findings quantitatively and raises the possibility that aggregate performance is overstated. The recent emergence of AI-specific reporting guidance is an encouraging corrective, but its influence on the primary literature is necessarily gradual, and much of the foundational corpus predates it (Liu et al., 2020).

A third gap is the uneven treatment of the challenges surveyed above. Data scarcity and, increasingly, generalisation are addressed with concrete technical strategies, whereas interpretability and bias are more often acknowledged than resolved. The dispute over whether post-hoc explanation is adequate (Selvaraju et al., 2017; Rudin, 2019) illustrates an unsettled methodological debate rather than a consolidated best practice. Similarly, fairness auditing and subgroup reporting are frequently recommended but inconsistently implemented in the primary studies reviewed, and the hidden-stratification findings suggest that headline results may conceal clinically consequential failures (Oakden-Rayner et al., 2020). Robustness to adversarial and naturally occurring perturbations is likewise studied more as a vulnerability to be demonstrated than as a property to be routinely certified (Finlayson et al., 2019).

Finally, the literature is dominated by a small number of data-rich domains, notably radiology, pathology, ophthalmology and dermatology, which benefit from standardised, high-volume imaging. Less is known about domains where data are scarcer or more heterogeneous, and about how models behave under the continual distribution shifts of real clinical workflows, where scanners are upgraded, protocols change and patient populations drift over time. These observations point towards prospective external validation, standardised reporting, integrated interpretability, systematic fairness evaluation, and post-deployment monitoring within a clear regulatory framework as the most pressing directions for future work.

Conclusion

The literature reviewed here describes a field that has moved with unusual speed from architectural innovation to clinically oriented demonstration. Convolutional networks, adapted through transfer learning and segmentation-specific designs and now supplemented by attention-based and self-supervised approaches, underpin a substantial research corpus spanning radiology, pathology, ophthalmology and dermatology (Litjens et al., 2017; Ronneberger, Fischer and Brox, 2015; Dosovitskiy et al., 2021). The evidence base, anchored in benchmark and clinician-comparison studies, indicates that deep learning can perform diagnostic image-interpretation tasks to a standard that warrants serious clinical interest (Gulshan et al., 2016; Esteva et al., 2017). At the same time, the synthesis makes clear that persistent challenges of data availability, generalisation, interpretability, bias and regulation temper these results, and that methodological weaknesses in study design limit confident inference about real-world utility (Topol, 2019; Kelly et al., 2019; Liu et al., 2020). The overarching conclusion drawn from the literature is therefore one of qualified optimism: the technical foundations are strong, but the translation of demonstrated performance into dependable, equitable and properly regulated clinical practice remains an open and actively contested problem.

References

Ardila, D., Kiraly, A.P., Bharadwaj, S., Choi, B., Reicher, J.J., Peng, L., Tse, D., Etemadi, M., Ye, W., Corrado, G., Naidich, D.P. and Shetty, S. (2019) ‘End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography’, Nature Medicine, 25(6), pp. 954-961.

Azizi, S., Mustafa, B., Ryan, F., Beaver, Z., Freyberg, J., Deaton, J., Loh, A., Karthikesalingam, A., Kornblith, S., Chen, T., Natarajan, V. and Norouzi, M. (2021) ‘Big self-supervised models advance medical image classification’, in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, pp. 3478-3488.

Benjamens, S., Dhunnoo, P. and Mesko, B. (2020) ‘The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database’, npj Digital Medicine, 3, 118.

Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S. and Fuchs, T.J. (2019) ‘Clinical-grade computational pathology using weakly supervised deep learning on whole slide images’, Nature Medicine, 25(8), pp. 1301-1309.

Coudray, N., Ocampo, P.S., Sakellaropoulos, T., Narula, N., Snuderl, M., Fenyo, D., Moreira, A.L., Razavian, N. and Tsirigos, A. (2018) ‘Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning’, Nature Medicine, 24(10), pp. 1559-1567.

De Fauw, J., Ledsam, J.R., Romera-Paredes, B., Nikolov, S., Tomasev, N., Blackwell, S., Askham, H., Glorot, X., O’Donoghue, B., Visentin, D. et al. (2018) ‘Clinically applicable deep learning for diagnosis and referral in retinal disease’, Nature Medicine, 24(9), pp. 1342-1350.

DeGrave, A.J., Janizek, J.D. and Lee, S.-I. (2021) ‘AI for radiographic COVID-19 detection selects shortcuts over signal’, Nature Machine Intelligence, 3(7), pp. 610-619.

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J. and Houlsby, N. (2021) ‘An image is worth 16×16 words: transformers for image recognition at scale’, in International Conference on Learning Representations (ICLR).

Ehteshami Bejnordi, B., Veta, M., Johannes van Diest, P., van Ginneken, B., Karssemeijer, N., Litjens, G., van der Laak, J.A.W.M. and the CAMELYON16 Consortium (2017) ‘Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer’, JAMA, 318(22), pp. 2199-2210.

Esteva, A., Kuprel, B., Novoa, R.A., Ko, J., Swetter, S.M., Blau, H.M. and Thrun, S. (2017) ‘Dermatologist-level classification of skin cancer with deep neural networks’, Nature, 542(7639), pp. 115-118.

Finlayson, S.G., Bowers, J.D., Ito, J., Zittrain, J.L., Beam, A.L. and Kohane, I.S. (2019) ‘Adversarial attacks on medical machine learning’, Science, 363(6433), pp. 1287-1289.

Goodfellow, I., Bengio, Y. and Courville, A. (2016) Deep Learning. Cambridge, MA: MIT Press.

Gulshan, V., Peng, L., Coram, M., Stumpe, M.C., Wu, D., Narayanaswamy, A., Venugopalan, S., Widner, K., Madams, T., Cuadros, J. et al. (2016) ‘Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs’, JAMA, 316(22), pp. 2402-2410.

He, K., Zhang, X., Ren, S. and Sun, J. (2016) ‘Deep residual learning for image recognition’, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, pp. 770-778.

Huang, G., Liu, Z., van der Maaten, L. and Weinberger, K.Q. (2017) ‘Densely connected convolutional networks’, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI, pp. 4700-4708.

Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J. and Maier-Hein, K.H. (2021) ‘nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation’, Nature Methods, 18(2), pp. 203-211.

Kelly, C.J., Karthikesalingam, A., Sudlow, M., Hughes, C., Rees, G. and King, D. (2019) ‘Key challenges for delivering clinical impact with artificial intelligence’, BMC Medicine, 17(1), 195.

Krizhevsky, A., Sutskever, I. and Hinton, G.E. (2012) ‘ImageNet classification with deep convolutional neural networks’, in Advances in Neural Information Processing Systems (NIPS) 25. Lake Tahoe, NV, pp. 1097-1105.

LeCun, Y., Bengio, Y. and Hinton, G. (2015) ‘Deep learning’, Nature, 521(7553), pp. 436-444.

Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., van der Laak, J.A.W.M., van Ginneken, B. and Sanchez, C.I. (2017) ‘A survey on deep learning in medical image analysis’, Medical Image Analysis, 42, pp. 60-88.

Liu, X., Faes, L., Kale, A.U., Wagner, S.K., Fu, D.J., Bruynseels, A., Mahendiran, T., Moraes, G., Shamdas, M., Kern, C. 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), pp. e271-e297.

Liu, X., Cruz Rivera, S., Moher, D., Calvert, M.J. and Denniston, A.K. (2020) ‘Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension’, Nature Medicine, 26(9), pp. 1364-1374.

McKinney, S.M., Sieniek, M., Godbole, V., Godwin, J., Antropova, N., Ashrafian, H., Back, T., Chesus, M., Corrado, G.S., Darzi, A. et al. (2020) ‘International evaluation of an AI system for breast cancer screening’, Nature, 577(7788), pp. 89-94.

Oakden-Rayner, L., Dunnmon, J., Carneiro, G. and Re, C. (2020) ‘Hidden stratification causes clinically meaningful failures in machine learning for medical imaging’, in Proceedings of the ACM Conference on Health, Inference, and Learning (CHIL). Toronto, pp. 151-159.

Obermeyer, Z., Powers, B., Vogeli, C. and Mullainathan, S. (2019) ‘Dissecting racial bias in an algorithm used to manage the health of populations’, Science, 366(6464), pp. 447-453.

Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D., Bagul, A., Langlotz, C., Shpanskaya, K., Lungren, M.P. and Ng, A.Y. (2017) ‘CheXNet: radiologist-level pneumonia detection on chest X-rays with deep learning’, arXiv preprint arXiv:1711.05225.

Rieke, N., Hancox, J., Li, W., Milletari, F., Roth, H.R., Albarqouni, S., Bakas, S., Galtier, M.N., Landman, B.A., Maier-Hein, K. et al. (2020) ‘The future of digital health with federated learning’, npj Digital Medicine, 3, 119.

Ronneberger, O., Fischer, P. and Brox, T. (2015) ‘U-Net: convolutional networks for biomedical image segmentation’, in Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2015. Munich, pp. 234-241.

Rudin, C. (2019) ‘Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead’, Nature Machine Intelligence, 1(5), pp. 206-215.

Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D. and Batra, D. (2017) ‘Grad-CAM: visual explanations from deep networks via gradient-based localization’, in Proceedings of the IEEE International Conference on Computer Vision (ICCV). Venice, pp. 618-626.

Shen, D., Wu, G. and Suk, H.-I. (2017) ‘Deep learning in medical image analysis’, Annual Review of Biomedical Engineering, 19, pp. 221-248.

Simonyan, K. and Zisserman, A. (2015) ‘Very deep convolutional networks for large-scale image recognition’, in International Conference on Learning Representations (ICLR). San Diego, CA.

Ting, D.S.W., Cheung, C.Y.-L., Lim, G., Tan, G.S.W., Quang, N.D., Gan, A., Hamzah, H., Garcia-Franco, R., San Yeo, I.Y., Lee, S.Y. et al. (2017) ‘Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes’, JAMA, 318(22), pp. 2211-2223.

Topol, E.J. (2019) ‘High-performance medicine: the convergence of human and artificial intelligence’, Nature Medicine, 25(1), pp. 44-56.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L. and Polosukhin, I. (2017) ‘Attention is all you need’, in Advances in Neural Information Processing Systems (NIPS) 30. Long Beach, CA, pp. 5998-6008.

Willemink, M.J., Koszek, W.A., Hardell, C., Wu, J., Fleischmann, D., Harvey, H., Folio, L.R., Summers, R.M., Rubin, D.L. and Lungren, M.P. (2020) ‘Preparing medical imaging data for machine learning’, Radiology, 295(1), pp. 4-15.

Zech, J.R., Badgeley, M.A., Liu, M., Costa, A.B., Titano, J.J. and Oermann, E.K. (2018) ‘Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study’, PLoS Medicine, 15(11), e1002683.