Clisp: A Robust Interactive Segmentation Framework for Pathological Images.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Computational Pathology

Background:

• Deep learning models excel at pathological image segmentation but require extensive annotated data, hindering clinical application.
• Interactive segmentation methods reduce annotation effort but often lack generalization due to small-scale training datasets.

Purpose of the Study:

• To introduce Clisp, a robust click-based interactive segmentation framework for efficient and low-cost annotation of pathological images.
• To enhance the generalization capabilities of interactive segmentation models for diverse pathological imaging tasks.

Main Methods:

• Developed a click-based interactive segmentation framework (Clisp) utilizing a heavy-parameter vision transformer for feature extraction.
• Constructed a large-scale, multi-source pathological image dataset comprising approximately 79,000 images for supervised training.
• Evaluated Clisp's performance and generalization ability across 7 open-source datasets.

Main Results:

• Clisp demonstrated superior segmentation performance compared to baseline methods on the majority of evaluated datasets.
• The proposed framework exhibited outstanding generalization capabilities across different segmentation tasks and objects.
• The use of a large-scale dataset and a powerful vision transformer encoder contributed to improved performance.

Conclusions:

• Clisp offers a practical solution for reducing the annotation burden in pathological image segmentation.
• The framework's strong generalization ability makes it suitable for a wide range of clinical applications.
• This work highlights the potential of interactive, deep learning-based approaches for efficient medical image analysis.