Unit-Based Histopathology Tissue Segmentation via Multi-Level Feature Representation

2025-07-16

AI Summary

Addressing the challenge of jointly optimizing accuracy, efficiency, and annotation cost in histopathological tissue segmentation, this paper proposes the Unitized Tile-based Segmentation (UTS) framework. UTS employs 32×32-pixel tiles as atomic classification units—eliminating expensive pixel-level annotations and substantially reducing both manual annotation burden and computational overhead. To capture discriminative morphological details and global tissue context, we introduce a novel multi-level Vision Transformer (L-ViT) architecture with hierarchical feature fusion. Evaluated on a large-scale H&E-stained dataset comprising 459 regions and 386,371 tiles, UTS achieves superior performance over U-Net variants and state-of-the-art ViT-based baselines in three-class breast tissue segmentation (tumor/stroma/other). The framework enables robust tumor–stroma quantification and surgical margin assessment, demonstrating strong clinical applicability.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Security and Privacy: Data transparency and provenance

Abstract

We propose UTS, a unit-based tissue segmentation framework for histopathology that classifies each fixed-size 32 * 32 tile, rather than each pixel, as the segmentation unit. This approach reduces annotation effort and improves computational efficiency without compromising accuracy. To implement this approach, we introduce a Multi-Level Vision Transformer (L-ViT), which benefits the multi-level feature representation to capture both fine-grained morphology and global tissue context. Trained to segment breast tissue into three categories (infiltrating tumor, non-neoplastic stroma, and fat), UTS supports clinically relevant tasks such as tumor-stroma quantification and surgical margin assessment. Evaluated on 386,371 tiles from 459 H&E-stained regions, it outperforms U-Net variants and transformer-based baselines. Code and Dataset will be available at GitHub.

Problem

• Segment histopathology tissue efficiently using fixed-size tiles
• Reduce annotation effort while maintaining segmentation accuracy
• Classify breast tissue into clinically relevant categories

Innovation

• Unit-based segmentation with 32x32 tiles
• Multi-Level Vision Transformer (L-ViT) for features
• Outperforms U-Net and transformer baselines