Deep Active Self-paced Learning for Biomedical Image Analysis

Abstract

Automatic and accurate analysis in biomedical images (e.g., image classification, lesion detection and segmentation) plays an important role in computer-aided diagnosis of common human diseases. However, this task is challenging due to the need of sufficient training data with high quality annotation, which is both time-consuming and costly to obtain. In this chapter, we propose a novel Deep Active Self-paced Learning (DASL) strategy to reduce annotation effort and also make use of unannotated samples, based on a combination of Active Learning (AL) and Self-Paced Learning (SPL) strategies. To evaluate the performance of the DASL strategy, we apply it to two typical problems in biomedical image analysis, pulmonary nodule segmentation in 3D CT images and diabetic retinopathy (DR) identification in digital retinal fundus images. In each scenario, we propose a novel deep learning model and train it with the DASL strategy. Experimental results show that the proposed models trained with our DASL strategy perform much better than those trained without DASL using the same amount of annotated samples.