MG-Net

MG-Net enhances thymoma segmentation in preoperative contrast-enhanced computed tomography (CECT) images by integrating multi-level feature interactions and global-awareness mechanisms to improve encoder-decoder representations.

Key Features:

• Enhanced Global-Awareness: Incorporates a multi-level feature interaction and integration framework including a Cross-Attention Block (CAB) for pixel-wise cross-level interactions that produces a Global Enhanced Convolution Block, a Global Spatial Attention Module that fuses coarse- and fine-grain information for semantic consistency, and an Adaptive Attention Fusion Module in the decoder for multi-scale feature aggregation.
• Superior Segmentation Performance: Demonstrates improved segmentation accuracy compared to state-of-the-art models on self-collected CECT datasets and the NIH Pancreas-CT dataset.
• Generalization Ability: Qualitative and quantitative experiments indicate robust generalization across different segmentation tasks beyond thymomas.

Scientific Applications:

• Thymoma segmentation: Automatic delineation of thymomas from CECT images to support preoperative diagnosis and surgical planning through accurate tumor boundary identification.
• Oncological imaging research and cross-dataset evaluation: Evaluation and development of segmentation methods across self-collected CECT datasets and the NIH Pancreas-CT dataset to study generalization in medical image segmentation.

Methodology:

Implements a deep learning encoder-decoder architecture with multi-level feature interaction via a Cross-Attention Block (CAB) that computes pixel-wise cross-level interactions to form a Global Enhanced Convolution Block, a Global Spatial Attention Module for coarse- and fine-grain fusion, and an Adaptive Attention Fusion Module in the decoder for adaptive multi-scale feature aggregation.