Medical artificial intelligence (AI) offers great potential for recognizing signs of health conditions in retinal images and expediting the diagnosis of eye diseases and systemic disorders¹. However, the development of AI models requires substantial annotation and models are usually task-specific with limited generalizability to different clinical applications². Here, we present RETFound, a foundation model for retinal images that learns generalizable representations from unlabelled retinal images and provides a basis for label-efficient model adaptation in several applications. Specifically, RETFound is trained on 1.6 million unlabelled retinal images by means of self-supervised learning and then adapted to disease detection tasks with explicit labels. We show that adapted RETFound consistently outperforms several comparison models in the diagnosis and prognosis of sight-threatening eye diseases, as well as incident prediction of complex systemic disorders such as heart failure and myocardial infarction with fewer labelled data. RETFound provides a generalizable solution to improve model performance and alleviate the annotation workload of experts to enable broad clinical AI applications from retinal imaging.

The Epiretinal Membrane (ERM) is an ocular disease that can cause visual distortions and irreversible vision loss. Patient sight preservation relies on an early diagnosis and on determining the location of the ERM in order to be treated and potentially removed. In this context, the visual inspection of the images in order to screen for ERM signs is a costly and subjective process. In this work, we propose and study three end-to-end fully-automatic approaches for the simultaneous segmentation and screening of ERM signs in Optical Coherence Tomography images. These convolutional approaches exploit a multi-task learning context to leverage inter-task complementarity in order to guide the training process. The proposed architectures are combined with three different state of the art encoder architectures of reference in order to provide an exhaustive study of the suitability of each of the approaches for these tasks. Furthermore, these architectures work in an end-to-end manner, entailing a significant simplification of the development process since they are able to be trained directly from annotated images without the need for a series of purpose-specific steps. In terms of segmentation, the proposed models obtained a precision of 0.760 ± 0.050, a sensitivity of 0.768 ± 0.210 and a specificity of 0.945 ± 0.011. For the screening task, these models achieved a precision of 0.963 ± 0.068, a sensitivity of 0.816 ± 0.162 and a specificity of 0.983 ± 0.068. The obtained results show that these multi-task approaches are able to perform competitively with or even outperform single-task methods tailored for either the segmentation or the screening of the ERM. These results highlight the advantages of using complementary knowledge related to the segmentation and screening tasks in the diagnosis of this relevant pathology, constituting the first proposal to address the diagnosis of the ERM from a multi-task perspective.

Retinal Optical Coherence Tomography (OCT) allows the non-invasive direct observation of the central nervous system, enabling the measurement and extraction of biomarkers from neural tissue that can be helpful in the assessment of ocular, systemic and Neurological Disorders (ND). Deep learning models can be trained to segment the retinal layers for biomarker extraction. However, the onset of ND can have an impact on the neural tissue, which can lead to the degraded performance of models not exposed to images displaying signs of disease during training. We present a fully automatic approach for the retinal layer segmentation in multiple neurodegenerative disorder scenarios, using an annotated dataset of patients of the most prevalent NDs: Alzheimer's disease, Parkinson's disease, multiple sclerosis and essential tremor, along with healthy control patients. Furthermore, we present a two-part, comprehensive study on the effects of ND on the performance of these models. The results show that images of healthy patients may not be sufficient for the robust training of automated segmentation models intended for the analysis of ND patients, and that using images representative of different NDs can increase the model performance. These results indicate that the presence or absence of patients of ND in datasets should be taken into account when training deep learning models for retinal layer segmentation, and that the proposed approach can provide a valuable tool for the robust and reliable diagnosis in multiple scenarios of ND.

Epiretinal Membrane (ERM) is a disease caused by a thin layer of scar tissue that is formed on the surface of the retina. When this membrane appears over the macula, it can cause distorted or blurred vision. Although normally idiopathic, its presence can also be indicative of other pathologies such as diabetic macular edema or vitreous haemorrhage. ERM removal surgery can preserve more visual acuity the earlier it is performed. For this purpose, we present a fully automatic segmentation system that can help the clinicians to determine the ERM presence and location over the eye fundus using 3D Optical Coherence Tomography (OCT) volumes. The proposed system uses a convolutional neural network architecture to classify patches of the retina surface. All the 2D OCT slices of the 3D OCT volume of a patient are combined to produce an intuitive colour map over the 2D fundus reconstruction, providing a visual representation of the presence of ERM which therefore facilitates the diagnosis and treatment of this relevant eye disease. A total of 2.428 2D OCT slices obtained from 20 OCT 3D volumes was used in this work. To validate the designed methodology, several representative experiments were performed. We obtained satisfactory results with a Dice Coefficient of 0.826 ± 0.112 and a Jaccard Index of 0.714 ± 0.155, proving its applicability for diagnosis purposes. The proposed system also demonstrated its simplicity and competitive performance with respect to other state-of-the-art approaches.

Central Serous Chorioretinopathy (CSC) is a retinal disorder caused by the accumulation of fluid, resulting in vision distortion. The diagnosis of this disease is typically performed through Optical Coherence Tomography (OCT) imaging, which displays any fluid buildup between the retinal layers. Currently, these fluid regions are manually detected by visual inspection a time-consuming and subjective process that can be prone to errors. A series of six deep learning-based automatic segmentation architectural configurations of different levels of complexity were trained and compared in order to determine the best model intended for the automatic segmentation of CSC-related lesions in OCT images. The best performing models were then evaluated in an external validation study. Furthermore, an intra- and inter-expert analysis was conducted in order to compare the manual segmentation performed by expert ophthalmologists with the automatic segmentation provided by the models. Test results of the best performing configuration achieved a mean Dice of $$0.868 \mathop {\pm } 0.056$$0.868 ± 0.056 in the internal dataset. In the external validation set, these models achieved a level of agreement with human experts of up to 0.960 in terms of Kappa coefficient, contrasting with a value of 0.951 for agreement between human experts. Overall, the models reached a better agreement with either of the human experts than these experts with each other, suggesting that automatic segmentation models for the detection of CSC-related lesions in OCT imaging can be useful tools for assessing this disease, reducing the workload of manual inspection and leading to a more robust and objective diagnosis method.

Glaucoma is the leading cause of irreversible blindness worldwide. The diagnosis process for glaucoma involves the measurement of the thickness of retinal layers in order to track its degeneration. The elongated shape of highly myopic eyes can hinder this diagnosis process, since it affects the OCT scanning process, producing deformations that can mimic or mask the degeneration caused by glaucoma. In this work, we present the first comprehensive cross-disease analysis that is focused on the anatomical structures most impacted in glaucoma and high myopia patients, facilitating precise differential diagnosis from those solely afflicted by myopia. To achieve this, a fully automatic approach for the retinal layer segmentation was specifically tailored for the accurate measurement of retinal thickness in both highly myopic and emmetropic eyes. To the best of our knowledge, this is the first approach proposed for the analysis of retinal layers in circumpapillary optical coherence tomography images that takes into account the elongation of the eyes in myopia, thus addressing critical diagnostic needs. The results from this study indicate that the temporal superior (mean difference 11.1μm, p<0.05), nasal inferior (13.1μm, p<0.01) and temporal inferior (13.3μm, p<0.01) sectors of the retinal nerve fibre layer show the most significant reduction in retinal thickness in patients of glaucoma and myopia with regards to patients of myopia.