Stable classification of diabetic structures from incorrectly labeled optical coherence tomography angiography en face images using multi instance learning

Abstract

We present a multiple instance learning-based network, MIL-ResNet14, detecting biomarkers for diabetic retinopathy in a widefield optical coherence tomography angiography dataset with high accuracy, without the necessity of annotations other than the information of whether a scan stems from a diabetic patient or not. Previously introduced deep learning-based classifiers were able to support the detection of diabetic biomarkers in OCTA images, however, require expert labeling on a pixel-level, a labor-intensive and expensive process. We evaluated our proposed architecture against two proven-capable classifiers, ResNet14 and VGG16. The dataset we applied for this study was acquired with a MHz A-Scan rate widefield Swept Source-OCT device. We utilized a total of 352 en face images, displaying retinal vasculature over a field of view of 18 mm x 18 mm. MIL-ResNet14 outperformed both other networks with an F-score of 0.95, a precision of 0.909 and an Area Under the Curve of 0.973. In addition, we could show via saliency overlays of gradient-weighted class activation mappings onto the en face images, that MIL-ResNet14 pays special attention to clinically relevant biomarkers like ischemic areas and retinal vessel anomalies. This could therefore function as a vigorous diagnostic decision support tool for clinical ophthalmologic screenings.