RGSB-UNet: Hybrid Deep Learning Framework for Tumour Segmentation in Digital Pathology Images.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Colorectal cancer (CRC) presents significant global health challenges due to high incidence and mortality rates.
• Early detection through screening is crucial for improving patient outcomes.
• Manual analysis of whole slide images (WSIs) for CRC diagnosis is time-consuming and labor-intensive.

Purpose of the Study:

• To develop an automated deep learning framework for precise tumor segmentation in WSIs.
• To overcome the limitation of existing deep learning models in capturing global features for pathological image analysis.
• To improve the efficiency and accuracy of colorectal cancer screening.

Main Methods:

• Introduction of a hybrid deep learning framework, RGSB-UNet, utilizing a UNet architecture.
• Incorporation of residual ghost blocks with switchable normalization (RGS) and bottleneck transformer (BoT) for refined feature extraction.
• Implementation of class-wise dice loss (CDL) for effective network training.

Main Results:

• The RGSB-UNet framework demonstrates superior performance in extracting refined features and robustness across varying batch sizes.
• The model achieves state-of-the-art segmentation performance, particularly under small batch size conditions.
• Experimental validation on DigestPath2019 and GlaS datasets confirms superior evaluation scores and segmentation accuracy.

Conclusions:

• The proposed RGSB-UNet model offers a significant advancement in automated tumor segmentation for colorectal cancer detection.
• This hybrid deep learning approach effectively combines spatial-local correlations and long-distance feature dependencies for enhanced pathological image analysis.
• The RGSB-UNet framework holds promise for improving the accuracy and efficiency of colorectal cancer screening and diagnosis.