Cross-Scale Guidance Integration Transformer for Instance Segmentation in Pathology Images

Yung-Ming Kuo ¹, Jia-Chun Sheng ², Chen-Hsuan Lo ²

⁰Department of Electronic EngineeringNational Formosa University Yunlin County 632 Taiwan.

IEEE Open Journal of Engineering in Medicine and Biology +

|

July 14, 2025

View abstract on PubMed

Summary

Pathologists manually grade adenocarcinoma by reviewing images. A new transformer model accurately segments individual gland cells, aiding computer-assisted diagnosis and improving grading consistency.

Area Of Science

Digital pathology
Computer-aided diagnosis
Medical image analysis

Background

Accurate adenocarcinoma grading requires manual review of pathology images by pathologists.
Manual grading faces challenges in inter-observer and intra-observer reproducibility.
Instance segmentation of individual gland cells is crucial for automated grading but remains difficult.

Purpose Of The Study

To develop an automated method for gland cell instance segmentation to assist in adenocarcinoma grading.
To improve the accuracy and reproducibility of computer-assisted grading of adenocarcinoma.
To address the challenge of segmenting individual gland cells of varying sizes.

Main Methods

A novel cross-scale guidance integration transformer was proposed for gland cell instance segmentation.
The network integrates multi-scale features using a cross-scale guidance integration module.
A decoder with mask attention utilizes integrated features for improved segmentation.

Main Results

The proposed method achieved state-of-the-art performance on two public gland cell datasets.
It demonstrated superior accuracy in segmenting individual gland cells compared to existing deep learning methods.
The approach effectively handles variations in gland cell size and morphology.

Conclusions

The cross-scale guidance integration transformer provides accurate gland cell segmentation.
This method assists pathologists in computer-assisted grading of adenocarcinoma.
The approach enhances reproducibility and efficiency in digital pathology workflows.

Keywords:

Instance segmentation pathology image self-attention transformer