Clisp: A Robust Interactive Segmentation Framework for Pathological Images

Recently, automatic segmentation algorithms based on deep learning have achieved promising results on various pathological image segmentation tasks. However, the inherent data-hungry nature of these methods and the high cost of annotating pathological images make them difficult to apply to practical clinical tasks. To address this problem, many studies have explored interactive segmentation methods to reduce the cost of annotating pathological images, which replace pixel-by-pixel annotation operations with clicks or other user instructions. Despite their effectiveness, most of these methods only utilize small-scale data resulting in poor generalization ability for different tasks or segmentation objects. In this paper, we propose a robust Click-based Interactive Segmentation framework for low-cost and interactive annotation of Pathological images, named Clisp. To improve the generalization ability, we construct a multi-source pathological image annotation dataset including around 79,000 images for supervised training of the proposed method. Besides, we adopt a heavy-parameter vision transformer as the image encoder to learn feature representations for large-scale data. We evaluate the proposed method on 7 open-source datasets, and the experimental results show that Clisp is superior to the baseline method in segmentation performance on most datasets and has outstanding generalization ability.