GCSA-SegFormer: Transformer-Based Segmentation for Liver Tumor Pathological Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Pathological image analysis is critical for tumor diagnosis but is time-consuming and subjective.
• High-resolution pathological images pose challenges for efficient and accurate interpretation.
• Artificial intelligence (AI) and deep learning offer potential solutions to improve diagnostic speed and reliability.

Purpose of the Study:

• To develop a novel Global Channel Spatial Attention (GCSA) module to enhance feature representation in pathological images.
• To integrate the GCSA module into the SegFormer architecture, creating the GCSA-SegFormer network.
• To improve the accuracy and efficiency of AI-driven pathological image diagnostics.

Main Methods:

• Proposed a Global Channel Spatial Attention (GCSA) module combining channel attention, channel shuffling, and spatial attention.
• Integrated the GCSA module into the SegFormer deep learning architecture.
• Evaluated the GCSA-SegFormer network on a liver dataset and the ICIAR 2018 BACH dataset.

Main Results:

• The GCSA-SegFormer achieved a 1.12% increase in Mean Intersection over Union (MIoU) and a 1.15% increase in Mean Pixel Accuracy (MPA) on the liver dataset.
• On the BACH dataset, the GCSA-SegFormer improved MIoU by 1.26% and MPA by 0.39% compared to baseline models.
• Demonstrated superior performance against seven other semantic segmentation methods.

Conclusions:

• The proposed GCSA module effectively enhances the representational capability of feature maps for pathological image analysis.
• The GCSA-SegFormer network accurately captures global and detailed features in complex pathological images.
• This AI-driven approach shows significant potential for improving the speed, accuracy, and reliability of tumor diagnosis.