Weakly-Supervised Lesion Segmentation with Self-Guidance by CT Intensity Clustering

To aid clinicians diagnose diseases and monitor lesion conditions more efficiently, automated lesion segmentation is a convincing approach. As it is time-consuming and costly to obtain pixel-level annotations, weakly-supervised learning has become a promising trend. Recent works based on Class Activation Mapping (CAM) achieve success for natural images, but they have not fully utilized the intensity property in medical images such that the performance may not be good enough. In this work, we propose a novel weakly-supervised lesion segmentation framework with self-guidance by CT intensity clustering. The proposed method takes full advantages of the properties that CT intensity represents the density of materials and partitions pixels into different groups by intensity clustering. Clusters with high lesion probability determined by the CAM are selected to generate lesion masks. Such lesion masks are used to derive self-guided loss functions which improve the CAM for better lesion segmentation. Our method achieves the Dice score of 0.5874 on the COVID-19 dataset and 0.4534 on the Liver Tumor Segmentation Challenge (LiTS) dataset.