Weakly supervised learning for retinal lesion detection

Abstract

Diabetic retinopathy (DR), a complication of long-term diabetes, is a common cause of visual impairment. Accurate detection of the lesions in retinal images is crucial for its opportune prevention in clinical practice. However, this task is time-consuming for ophthalmologists, so there is a need for automatic methods. Furthermore, such methods are typically trained in weakly supervised way, to compensate for small amount of full annotated data. The purpose of this work is to develop an effective algorithm for detecting retinal lesions, when only image-level labels and coarse masks are available, with minimal manual feature extraction. The dataset under consideration is DiaRetDB1, however, the IDRiD dataset is also utilised for comparison and algorithms' evaluation. The metrics for 3 types of lesions are reported, in particular the ROC AUC of 0.86 and PR AUC of 0.08 at pixel-level are reported for microaneurysms.