Sequence of Morphological Changes Preceding Atrophy in Intermediate AMD Using Deep Learning.

Purpose: Investigating the sequence of morphological changes preceding outer plexiform layer (OPL) subsidence, a marker preceding geographic atrophy, in intermediate AMD (iAMD) using high-precision artificial intelligence (AI) quantifications on optical coherence tomography imaging. Methods: In this longitudinal observational study, individuals with bilateral iAMD participating in a multicenter clinical trial were screened for OPL subsidence and RPE and outer retinal atrophy. OPL subsidence was segmented on an A-scan basis in optical coherence tomography volumes, obtained 6-monthly with 36 months follow-up. AI-based quantification of photoreceptor (PR) and outer nuclear layer (ONL) thickness, drusen height and choroidal hypertransmission (HT) was performed. Changes were compared between topographic areas of OPL subsidence (AS), drusen (AD), and reference (AR). Results: Of 280 eyes of 140 individuals, OPL subsidence occurred in 53 eyes from 43 individuals. Thirty-six eyes developed RPE and outer retinal atrophy subsequently. In the cohort of 53 eyes showing OPL subsidence, PR and ONL thicknesses were significantly decreased in AS compared with AD and AR 12 and 18 months before OPL subsidence occurred, respectively (PR: 20 µm vs. 23 µm and 27 µm [P < 0.009]; ONL, 84 µm vs. 94 µm and 98 µm [P < 0.008]). Accelerated thinning of PR (0.6 µm/month; P < 0.001) and ONL (0.8 µm/month; P < 0.001) was observed in AS compared with AD and AR. Concomitant drusen regression and hypertransmission increase at the occurrence of OPL subsidence underline the atrophic progress in areas affected by OPL subsidence. Conclusions: PR and ONL thinning are early subclinical features associated with subsequent OPL subsidence, an indicator of progression toward geographic atrophy. AI algorithms are able to predict and quantify morphological precursors of iAMD conversion and allow personalized risk stratification.

3DTINC: Time-Equivariant Non-Contrastive Learning for Predicting Disease Progression from Longitudinal OCTs.

Self-supervised learning (SSL) has emerged as a powerful technique for improving the efficiency and effectiveness of deep learning models. Contrastive methods are a prominent family of SSL that extract similar representations of two augmented views of an image while pushing away others in the representation space as negatives. However, the state-of-the-art contrastive methods require large batch sizes and augmentations designed for natural images that are impractical for 3D medical images. To address these limitations, we propose a new longitudinal SSL method, 3DTINC, based on non-contrastive learning. It is designed to learn perturbation-invariant features for 3D optical coherence tomography (OCT) volumes, using augmentations specifically designed for OCT. We introduce a new non-contrastive similarity loss term that learns temporal information implicitly from intra-patient scans acquired at different times. Our experiments show that this temporal information is crucial for predicting progression of retinal diseases, such as age-related macular degeneration (AMD). After pretraining with 3DTINC, we evaluated the learned representations and the prognostic models on two large-scale longitudinal datasets of retinal OCTs where we predict the conversion to wet-AMD within a six-month interval. Our results demonstrate that each component of our contributions is crucial for learning meaningful representations useful in predicting disease progression from longitudinal volumetric scans.

Deep survival modeling of longitudinal retinal OCT volumes for predicting the onset of atrophy in patients with intermediate AMD.

In patients with age-related macular degeneration (AMD), the risk of progression to late stages is highly heterogeneous, and the prognostic imaging biomarkers remain unclear. We propose a deep survival model to predict the progression towards the late atrophic stage of AMD. The model combines the advantages of survival modelling, accounting for time-to-event and censoring, and the advantages of deep learning, generating prediction from raw 3D OCT scans, without the need for extracting a predefined set of quantitative biomarkers. We demonstrate, in an extensive set of evaluations, based on two large longitudinal datasets with 231 eyes from 121 patients for internal evaluation, and 280 eyes from 140 patients for the external evaluation, that this model improves the risk estimation performance over standard deep learning classification models.