Automatic multi-tissue segmentation in pancreatic pathological images with selected multi-scale attention network.

Area of Science:

• Pathology
• Medical Imaging
• Computer Vision

Background:

• Pathological image analysis is crucial for assessing pancreatic ductal adenocarcinoma (PDAC) malignancy.
• Accurate segmentation of tumor cells and surrounding tissues is vital for morphological analysis but challenging due to appearance variations.

Purpose of the Study:

• To propose a novel deep learning model, SMANet, for segmenting multiple tissue types in PDAC pathological images.
• To enhance the accuracy of segmentation for tumor cells, blood vessels, nerves, islets, and ducts.

Main Methods:

• Developed a Selected Multi-scale Attention Network (SMANet) incorporating Selection Unit (SU) and Multi-scale Attention (MA) modules.
• The MA module uses spatial and channel attention to refine features across different scales.
• An original-feature fusion unit was introduced to improve segmentation of small structures like islets and ducts.

Main Results:

• SMANet demonstrated superior performance compared to state-of-the-art deep learning methods on PDAC pathological images.
• Achieved competitive results on the GlaS challenge dataset.
• Attained mDice of 0.769 and mIoU of 0.665 on the PDAC dataset.

Conclusions:

• SMANet effectively enhances relevant information and suppresses noise, leading to improved segmentation accuracy.
• The model's ability to integrate multi-scale features and original image information addresses segmentation challenges in PDAC images.
• The proposed method shows significant potential for improving diagnostic accuracy in pancreatic cancer pathology.