Semi-Supervised Learning Improves Model Performance for Retinal Vessel Segmentation on Infrared Reflectance Imaging

Abstract

Purpose : A significant challenge in training deep learning models for a semantic segmentation task is the limited availability of labeled training data. Semi-supervised learning uses both unlabeled and labeled training data and has been shown to improve model performance compared to the traditional paradigm of using only labeled data. We demonstrate the ability of a semi-supervised framework to improve performance on a multiclass retinal vessel segmentation task from infrared reflectance (IR) images of the retina.

Methods : We used all 22 images with manually labeled ground truth segmentation masks of the arteries and veins from the publicly available RAVIR dataset with a 19-3 training-validation split. All images were disc centered IR images acquired with a Heidelberg Spectralis camera with a 30o field of view. We also used 1127 unlabeled images from an internal dataset when training the semi-supervised model. These images were IR images acquired as part of the regular macular OCT acquisition protocol on the Heidelberg Spectralis.

We trained DeepLabv3 with a Resnet50 backbone as our baseline model and used Cross Pseudo Supervision (CPS) with DeepLabv3 with Resnet50 as our semi-supervised framework. Both baseline and semi-supervised models saw a similar number of iterations of the labeled data during training and had the same hyperparameters. The metric reported was mean intersection-over-union (IOU) for the artery and vein classes calculated on a separate 19 image test set provided in the RAVIR dataset.

Results : The mean IOU for the baseline and supervised models on the test set were 43.77 (95% CI: 26.7 - 57.9) and 57.02 (95% CI: 43.8-70.3). Semi-supervised learning was significantly better than baseline (p <0.01) and improved mean IOU by 30.27%. Model size for inference was the same for both methods.

Conclusions : Semi-supervised learning improves model performance without increasing the number of labeled training images in the dataset. This method may be used to improve model performance for a semantic segmentation task when there is an abundance of unlabeled data.

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

 

Mean intersection over union (IOU) metric on the test set for the baseline model compared to the Cross Pseudo Supervision (CPS) model. The error bars represent the 95% CI.

 

Qualitative representation of output predictions for the cross pseudo supervision (CPS) model and baseline model.