Contrastive Learning Driven Self-Supervised Framework for Segmentation of Biomarker of Diabetic Macular Edema

Abstract

Purpose : To develop a self-supervised model based on deep learning (DL) and contrastive learning (CL) for segmentation of diabetic macular edema (DME)-induced lesions based on OCT scans.

Methods : We developed a self-supervised DL-CL model with two stages of training (Fig 1) based on 13,000 OCT scans collected from the publicly available Zang dataset (dataset 1) and 610 OCT scans collected from the publicly available Duke-II dataset (dataset 2). In the first stage, we developed an encoder-decoder model based on RAG-NET architecture and employed a cross entropy loss function (L_ic) then trained the model based on annotated OCT scans with manifestations of intra-retinal fluid (IRF), sub-retinal fluid (SRF), and hard exudates (HE) lesions (from the first dataset). In the second stage, we developed a CL model and employed a contrastive loss function (L_c), to learn retinal lesion segmentation in a self-supervised manner based on unlabeled OCT scans (from the second dataset). This is achieved by imposing self-supervised constraining on the segmentation using contrastive loss function. We trained our model based on a subset of OCT scans from both datasets and validated the model based on an independent subset from dataset 1 with 3000 OCT scans (Fig 2).

Results : The area under the receiver operating characteristics curves (AUCs) of the model in detecting DME lesion for labeled and unlabeled data were 0.96 and 0.94 respectively.

Conclusions : We developed a self-supervised contrastive learning-driven model that segments DME lesions without any labels. The proposed model may help clinicals for diagnosing DME.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Diagram of the proposed self-supervised model for detecting DME. Training phase 1: supervised learning based on encoder-decoder network using RAG-NET. Training phase 2: contrastive learning model learns the segmentation of retinal lesions based on previous learned experience and prior knowledge.

 

Qualitative and quantitative evaluations. (A) Segmentation results of the proposed model for identification of DME-induced lesions (IRF in red, SRF in yellow, and HE in blue). Columns (a) and (b) present normal and DME OCT images, respectively. Column (c) represents the ground truth segmentation of scans in the previous column. Column (d) shows the model’s segmented outcome of normal images. Column (e) presents the identified lesions by the model. (B) AUC of the model based on labeled and unlabeled images.