Lesion-aware knowledge distillation for diabetic retinopathy lesion segmentation

Retinal fundus images have been widely utilized for screening Diabetic Retinopathy (DR). The lesion information contained in these images is indispensable for the diagnosis of DR. The acquisition of lesion information depends on the sophisticated lesion segmentation methods. Nevertheless, the existing lesion segmentation methods highly rely on huge computational complexity and massive storage, making it difficult to apply in real-world clinical scenarios. Knowledge distillation (KD) has become an essential tool to reduce the computational complexity of the network. However, the lesion regions being insignificant in fundus images, directly applying the current KD methods cannot adequately transfer sufficient lesion knowledge, which restricts the learning of knowledge distillation. In essence, the challenge is how to enhance the focus of the KD process on lesion regions and to transfer more comprehensive pathological knowledge to the student network. Considering the importance of lesion regions in fundus images and the global semantic relations among lesion regions across various fundus images, we propose a Lesion-aware Knowledge Distillation (LKD) framework focusing on the transfer of lesion knowledge. The key contribution of the proposed framework is the creation of lesion embedding queue from the global training samples, which facilitates the transfer of global pathology knowledge from the teacher network to the student network, thus promoting the acquisition of lesion-related knowledge. Furthermore, we propose a self-paced hard sample learning strategy for knowledge transfer of lesion embedding queue, which additionally improves the efficiency of knowledge transfer. We evaluate LKD on IDRiD and DDR benchmark datasets, the overall performance of the proposed mehtod improves 2.1% AUPR as well as 2.2% DICE and 1.5% AUPR as well as 2.2% DICE compared to the previous best results. A particular improvement of 2.3% and 3.2% in DICE are achieved on the IDRiD dataset for tiny lesions, i.e., MA and HE, respectively. Our code is available at https://github.com/YaqiWangCV/LKD.