Cross-Scale Fusion Transformer for Histopathological Image Classification.

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Histopathological images are crucial for disease diagnosis, but pathologist availability and image variability pose challenges.
• Existing methods struggle with diverse image scales and magnifications, hindering generalizable solutions.
• Automated histopathological image classification is needed to support pathologists and improve diagnostic efficiency.

Purpose of the Study:

• To propose a novel Cross-Scale Fusion (CSF) transformer for robust histopathological image classification.
• To develop a method that effectively integrates multi-field-of-view patch embeddings.
• To address the limitations of existing methods in handling variations in organs, cell sizes, and magnification factors.

Main Methods:

• The proposed CSF transformer utilizes a multiple field-of-view patch embedding module.
• It incorporates transformer encoders and novel cross-fusion modules for integrating information.
• The architecture learns cross-scale contextual correlations from different patch embeddings.

Main Results:

• The CSF transformer demonstrated superior performance compared to state-of-the-art methods on four public datasets.
• It effectively integrates patch embeddings from different fields-of-view, capturing cross-scale contextual correlations.
• The method achieved better results than convolutional neural network-based and other transformer-based approaches.

Conclusions:

• The CSF transformer offers an effective and generalizable solution for histopathological image classification.
• It addresses challenges related to image variability and pathologist workload with improved efficiency.
• The proposed architecture shows promise for advancing automated disease diagnosis in digital pathology.