Navigating Through Whole Slide Images With Hierarchy, Multi-Object, and Multi-Scale Data.

Area of Science:

• Computational Pathology
• Deep Learning
• Medical Image Analysis

Background:

• Segmenting whole slide images (WSIs) with deep learning models using few training samples is challenging due to similar morphological structures.
• Pathologists utilize hierarchical feature organization and multi-scale analysis for accurate tissue identification.

Purpose of the Study:

• To develop a vision model, the Navigator, that mimics pathologist multi-scale diagnostic workflows for WSI segmentation.
• To address the challenge of limited annotated data in computational pathology.

Main Methods:

• The Navigator model employs a multi-scale approach, searching low-resolution images and gradually zooming to localize finer microanatomical classes.
• A novel semi-supervised framework, S5CL v2, was utilized for training the Navigator on sparsely annotated samples.

Main Results:

• The Navigator effectively detects coarse-grained patterns at lower resolutions and fine-grained features at higher resolutions.
• The model achieved up to an 8% improvement in F1 score on various datasets, including TCGA-COAD-30CLS and Erlangen cohorts.

Conclusions:

• The Navigator model demonstrates a promising approach for efficient and accurate WSI segmentation, particularly in data-scarce scenarios.
• The multi-scale navigation strategy and semi-supervised learning framework offer significant advancements in computational pathology.