Stable classification of diabetic structures from incorrectly labeled OCTA enface images using multiple instance learning

Abstract

Purpose : Early diagnosis and monitoring of the progression of diabetic retinopathy (DR) are crucial to help prevent permanent loss of vision in patients. Artificial intelligence (AI)-algorithms have been utilized to classify DR and even its severity. However, most algorithms are based on supervised learning which requires precise labels for the training data. We minimize this necessity via a multiple instance learning (MIL) machine learning approach.

Methods : MIL only uses a set of instances, so-called bags, in which all instances inherit one global bag label - in our case diabetic or not. The algorithm then learns to classify these instances, based on the inherited bag label. 352 OCTA enface images from 102 patients (d) and 40 healthy (h) volunteers were collected and split - 211/64 (d/h) for training, 24/24 for validation 22/8 for testing. Bags are preprocessed and then run through the classifier (Figure 1). MIL-ResNet14 was evaluated against two other network architectures that have been proven capable: ResNet14, and VGG16.

Results : Our proposed network, MIL-ResNet14, provided an F1-score (harmonic mean of precision and recall) of 0.950, VGG16 of 0.857, and ResNet14 of 0.909. MIL-ResNet14 outperformed both ResNet14 and VGG16 in detecting vasculature changes of diabetes, such as ischemic areas and abnormal vessel branches, which we validate visually with Grad-CAM overlays (Figure 2). MIL-ResNet14 also accurately classified cases in-between healthy and very diabetic more accurately than both other networks.

Conclusions : Minimizing the necessity of expert grader labelling makes AI-based classification algorithms much easier in terms of data preparation. From our comparison with established classifiers, we conclude that MIL has a regularizing effect on incorrectly labeled data sets and is therefore a clinically more reliable classification architecture than previously proposed methods.

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

 

Figure 1: Architecture of the proposed network, MIL-ResNet14. Images are pre-processed and then sub-divided into instances, each inheriting the global image label for training and validation (a) and are then passed into the instance classifier and MIL pooling parts (b).

 

Figure 2: Grad-CAM overlays of all three networks onto OCTA enface example image of a diabetic patient; (a) ResNet14; (b): VGG16; (c): ResNet14. Blue arrows indicate ischemic areas and pink arrows point to vascular abnormalities.