Monitoring Disease Progression with Stable Diffusion in Retinopathy of Prematurity

Abstract

Purpose : Retinopathy of Prematurity (ROP) is the leading cause of childhood blindness in premature infants. While only 5-10% of total cases progress to severe ROP, it is quintessential to understand the physiology & progression of ROP, for early diagnosis & effective & timely treatment. In this study, we leverage a Stable Diffusion (SD) model to build a class-conditioned ROP image generation tool & further combine it with image editing techniques to enable easy transformation of disease severity for a given ROP image

Methods : We cast the problem of visualizing ROP progression as an image editing problem. Given an input image of any of the Normal/Pre-Plus/Plus class, the goal is to transform it into another class, while preserving the global structure & some local attributes so that the transformed image is still perceived as an image belonging to the same patient. We first fine-tune a Text-to-Image SD model on the iROP dataset (5500 images), thus allowing the model to generate realistic images of each class prompted by textual inputs. We next employ two image editing techniques: Null Text Inversion & Prompt-to-Prompt Image Editing, essentially allowing image modification through attention reweighting (Fig.1) during the forward diffusion process. To evaluate SD-based image generation, we calculate Frechet Inception Distance (FID) score between real & generated images as well as Plus disease classifier accuracy when tested on class-conditioned generated images. To evaluate image editing, we calculate correlation between ROP Vascular Severity Score (VSS) of a generated image & real image of increased severity for the same normal image.

Results : Our SD model achieves FID scores of 92, 114, and 137 for Normal, Pre-Plus, & Plus classes, respectively. We achieve 97% accuracy for generated plus class when tested with a Plus Classifier model. A correlation of 0.97 between ROP VSS of generated & real images indicates proportional attention weight adjustment enables the generation of required severity images

Conclusions : Our SD-based image editing tool allows easy visualization of the severity of the disease at a future date on new data samples, which may have no disease development yet or are at a moderate severity stage of the disease. This can provide clinicians with valuable insights into the potential progression of the disease & aid in early diagnosis & treatment, ultimately improving outcomes for premature infants.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.