An unsupervised method for histological image segmentation based on tissue cluster level graph cut.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning

Background:

• Deep learning models excel in medical image segmentation but require extensive histological annotations.
• Acquiring these annotations is labor-intensive and requires specialized expertise.
• Unsupervised methods are needed to reduce the annotation burden for supervised models.

Purpose of the Study:

• To develop an unsupervised method, TisCut, for segmenting histological images into meaningful compartments.
• To assist in generating annotations for downstream supervised segmentation models.
• To evaluate TisCut's performance in segmenting tumor and non-tumor regions.

Main Methods:

• TisCut employs a three-module approach: tissue object clustering, Voronoi diagram construction, and region adjacency graph integration.
• Clustering is based on spatial proximity and morphological features.
• The graph cut algorithm is used for image partitioning based on morphological features from the Voronoi diagram.

Main Results:

• TisCut achieved comparative performance against U-Net models in histological image segmentation.
• Jaccard index coefficients of approximately 70% and 85% were obtained for brain and skin histological images, respectively.
• The method demonstrated effectiveness in detecting necrosis and melanoma.

Conclusions:

• TisCut offers a viable unsupervised solution for histological image segmentation.
• The method shows potential for generating annotations when training data is scarce.
• TisCut can aid in the development of supervised models by reducing annotation effort.