TSFD-Net: Tissue specific feature distillation network for nuclei segmentation and classification.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Nuclei segmentation and classification in H&E-stained histology images is complex due to staining inconsistencies, clustered, and overlapping nuclei.
• Current methods using polygon representations or centroid distances have limitations in accuracy and robustness.

Purpose of the Study:

• To develop an advanced deep learning model for accurate nuclei segmentation and classification in histopathology images.
• To address challenges posed by image variability and complex cellular structures in digital pathology.

Main Methods:

• Proposed a novel Tissue-Specific Feature Distillation (TSFD) backbone to extract discriminative morphological features.
• Utilized a Bi-directional Feature Pyramid Network (BiFPN) for hierarchical feature representation and fusion.
• Introduced a combinational loss function for joint optimization and accelerated network convergence.

Main Results:

• The TSFD-Net significantly outperformed state-of-the-art methods including StarDist, Micro-Net, Mask-RCNN, Hover-Net, and CPP-Net on the PanNuke dataset.
• Achieved 50.4% mean and 63.77% binary panoptic quality on a dataset with 19 tissue types and 5 tumor classes.

Conclusions:

• TSFD-Net demonstrates superior performance in nuclei segmentation and classification for diverse histopathology images.
• The proposed method effectively leverages tissue-specific features and advanced network architecture for improved accuracy.