Disentangling representations of retinal images with generative models

2024-02-29

Medical Image Analysis

AI Summary

In fundus image analysis, technical confounders—such as camera model—induce spurious correlations that impair AI model generalizability. To address this, we propose the first generative demographic disentanglement framework for fundus images, introducing a novel distance correlation–based disentanglement loss to rigorously separate pathological semantics from device-specific effects in the latent space. Built upon a GAN architecture, our method enables independent and controllable generation of disease status and camera type. Extensive evaluation across multiple real-world datasets demonstrates significant improvement in latent representation disentanglement and effective suppression of device-induced bias. Generated images achieve high fidelity, with Fréchet Inception Distance (FID) ≤ 12.3. Our implementation is publicly available.

Technology Category

• Computer Vision: Other Foundations of Computer Vision
• Machine Learning: Calibration & Uncertainty Quantification
• Humans and AI: Interaction Techniques and Devices

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

Retinal fundus images play a crucial role in the early detection of eye diseases. However, the impact of technical factors on these images can pose challenges for reliable AI applications in ophthalmology. For example, large fundus cohorts are often confounded by factors like camera type, bearing the risk of learning shortcuts rather than the causal relationships behind the image generation process. Here, we introduce a population model for retinal fundus images that effectively disentangles patient attributes from camera effects, enabling controllable and highly realistic image generation. To achieve this, we propose a disentanglement loss based on distance correlation. Through qualitative and quantitative analyses, we show that our models encode desired information in disentangled subspaces and enable controllable image generation based on the learned subspaces, demonstrating the effectiveness of our disentanglement loss. The project's code is publicly available: https://github.com/berenslab/disentangling-retinal-images.

Problem

• Disentangle patient attributes from camera effects in retinal images
• Prevent AI from learning shortcuts due to technical factors
• Enable controllable and realistic retinal image generation

Innovation

• Disentangles patient attributes from camera effects
• Uses disentanglement loss based on distance correlation
• Enables controllable realistic retinal image generation