A Deep Learning Framework for Gastric Cancer Cell Segmentation with Multi-Scale Attention Mechanisms

Area of Science:

• Medical image analysis
• Computational pathology
• Deep learning in oncology

Background:

• Accurate segmentation of gastric cancer cells is crucial for early diagnosis and disease detection in pathology.
• Current segmentation methods face challenges including high annotation costs, indistinct lesion boundaries, and limited feature expression.

Purpose of the Study:

• To develop an advanced deep learning framework, MSAF-Net, for improved gastric cancer cell segmentation.
• To address limitations of existing methods by enhancing feature representation and boundary reconstruction.

Main Methods:

• Developed MSAF-Net, a novel U-Net based framework with architectural and loss function optimizations.
• Incorporated a Multi-scale Dilated Pooling Fusion Block in the encoder for enhanced multi-path interaction and feature diversity.
• Introduced a Dual-Channel Attention Block in the decoder for improved detail restoration and boundary reconstruction.
• Integrated a Diagonal Mahalanobis Consistency Loss to promote class compactness.

Main Results:

• MSAF-Net achieved a Dice score of 0.776 and an Accuracy of 0.821 on the SEED-Gastric Carcinoma Stage 1 dataset.
• The proposed method demonstrated superior performance compared to the baseline U-Net model.
• Results indicate significant improvements in feature expression, boundary sensitivity, and overall segmentation accuracy.

Conclusions:

• MSAF-Net effectively enhances gastric cancer cell segmentation accuracy and robustness.
• The novel architectural components and loss function contribute to precise quantification of cell morphology.
• This approach shows promise for improving early diagnosis and pathological analysis of gastric cancer.