Automatic Tumor Cellularity Measurement: AI-Based Pipeline for Multi-Organ Pathology Imaging

Area of Science:

• Digital pathology
• Computational imaging
• Artificial intelligence in oncology

Background:

• Tumor cellularity (TC) is critical for evaluating organ tumor burden but manual counting is impractical due to image volume and inter-observer variability.
• Existing AI models struggle with applying pancreas-trained data to colon images and accurately segmenting clustered cells.

Purpose of the Study:

• To develop and validate an AI pipeline for accurate and automated tumor cellularity assessment in digital pathology images.
• To address the challenges of cross-organ generalization and clustered cell segmentation in AI-driven pathology.

Main Methods:

• A novel pipeline incorporating channel normalization for consistent RGB standardization across organs.
• Introduction of CacoX, a specialized cell segmentation model utilizing Coordinate Attention Gates and non-local learning for precise localization.
• Application of a watershed algorithm to automatically separate clustered cells in segmentation outputs.

Main Results:

• The proposed pipeline achieved 3rd place in the PAIP 2023 Challenge.
• An Interclass Correlation (ICC) score of 95.69% demonstrates high agreement and accuracy.
• The method effectively handled cross-organ data challenges and improved clustered cell separation.

Conclusions:

• The developed AI pipeline offers a robust solution for automated tumor cellularity measurement in digital pathology.
• This approach enhances consistency, efficiency, and accuracy compared to manual methods.
• The findings highlight the potential of specialized AI models and algorithms for advancing cancer burden assessment.