Deep Learning Detection of Retinitis Pigmentosa Inheritance Forms through Synthetic Data Expansion of a Rare Disease Dataset

Abstract

Accurate classification of inheritance patterns is an integral part of diagnosis and genetic counseling for inherited retinal diseases (IRDs). Traditionally reliant on pedigree analysis, clinical phenotyping, and genetic testing, this process is often constrained by incomplete family history, ambiguous presentations, limited access to genetic testing, and inconclusive genetic test results. Deep learning (DL) applied to fundus imaging presents a promising approach for automated inference of inheritance modes; however, development has been hindered by the low prevalence of IRDs and the scarcity of annotated datasets. In this study, we focus on retinitis pigmentosa (RP), a highly heterogeneous disorder in both clinical presentation and genetic etiology. We present a first-in-class deep learning approach that leverages Vision Transformer (ViT) models to distinguish autosomal from X-linked RP using color fundus photography. To overcome challenges posed by limited data, we introduce an innovative variational autoencoder-based data expansion strategy, which improves inheritance pattern classification based on color fundus photos from 0.67 AUC to 0.79 AUC. Our findings demonstrate the potential of deep learning to uncover subtle phenotypic differences linked to genetic inheritance and introduce a novel training data augmentation method to render deep learning accessible to rare diseases.

Figure 1

Two-way combinatorial synthetic image generation using variational autoencoder.

Figure 2

Examples of Synthetic Images based on Autosomal Dominant (AD) RP images. Left to right A/B ratios: 0.1, 0.3, 0.5, 0.7, 0.9.

Figure 3

Workflow for Data Expansion and Vision Transformer Model Training. VAE = Variational Autoencoder, ViT = Vision Transformer.

Figure 4

Patients mean age (panel A) and median symptom duration at time of imaging (panel B). AD = Autosomal Dominant, AR = Autosomal Recessive, XL = X-linked Recessive and X-linked Carrier (combined for anonymization purposes). ** denotes p-value < 0.005 and *** denotes p-value < 0.0001

Figure 5

A. Receiver Operating Characteristic (ROC) curve with pooled AUC reported for RP ViT base (unexpanded) model. B. Confusion matrix for the base model.

Figure 6

A. Receiver Operating Characteristic (ROC) curves for RP-ViT base and Gen 1 (random noise) expanded models. B. Confusion matrix for Gen 1 (random noise)-expanded models.

Figure 7

Receiver Operating Characteristic (ROC) curves for RP-ViT base, Gen 1 random noise-expanded, and Gen 2 pair-wise combinatorial expanded models. B. Confusion matrix for Gen 2 RP-ViT model.