3D Multi-Attention Guided Multi-Task Learning Network for Automatic Gastric Tumor Segmentation and Lymph Node...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Radiology and diagnostic imaging

Background:

• Accurate gastric tumor segmentation and lymph node classification are crucial for clinical diagnosis and treatment planning.
• Current automated methods face challenges due to inhomogeneous intensity, ambiguous boundaries, and variable shapes in CT scans.

Purpose of the Study:

• To develop a novel 3D multi-attention guided multi-task learning network for simultaneous gastric tumor segmentation and lymph node classification.
• To address challenges in automated gastric cancer imaging analysis by leveraging complementary information across dimensions, scales, and tasks.

Main Methods:

• A 3D convolutional neural network incorporating scale-aware and task-aware attention mechanisms for refined multi-scale and task-specific feature learning.
• Utilized visual attention, adaptive spatial attention, and stage-wise deep supervision for shared feature learning.
• Employed segmentation-aware and classification-aware attention modules for task-specific feature enhancement.
• Balanced segmentation and classification tasks using combined loss functions with weight uncertainty.

Main Results:

• The proposed network achieved superior performance in gastric tumor segmentation and lymph node classification compared to state-of-the-art methods on an in-house CT dataset.
• Demonstrated promising results and potential for clinical application in gastric cancer diagnosis.
• Successfully extended to liver tumor segmentation, indicating good generalization capabilities.

Conclusions:

• The novel 3D multi-attention guided multi-task learning network effectively addresses the complexities of gastric tumor segmentation and lymph node classification.
• The method offers a robust and accurate solution for automated medical image analysis in oncology.
• The network's generalizability suggests its potential applicability to other medical segmentation tasks.