Weakly-Supervised TILs Segmentation Based on Point Annotations Using Transfer Learning with Point Detector and Projected-Boundary Regressor

In Whole Slide Image (WSI) analysis, detecting nuclei subtypes such as Tumor Infiltrating Lymphocytes (TILs) which are a primary bio-marker for cancer diagnosis, is an important yet challenging task. Though several conventional methods have been proposed and applied to target user's nuclei sub-types (e.g., TILs), they often fail to detect subtle differences between instances due to similar morphology across sub-types. To address this, we propose a novel decoupled segmentation architecture that leverages point annotations in a weakly-supervised manner to adapt to the nuclei sub-type. Our design consists of an encoder for feature extraction, a boundary regressor that learns prior knowledge from nuclei boundary masks, and a point detector that predicts the center positions of nuclei, respectively. Moreover, employing a frozen pre-trained nuclei segmenter facilitates easier adaptation to TILs segmentation via fine-tuning, while learning a decoupled point detector. To demonstrate the effectiveness of our approach, we evaluated on an in-house Melanoma TIL dataset, and report significant improvements over a state-of-the-art weakly-supervised TILs segmentation method, including conventional approaches based on pseudo-label construction.