A Hybrid U-Shaped Deep Learning Network for Intracerebral Hemorrhage Segmentation in CT Scans.
Authors
Affiliations (4)
Affiliations (4)
- School of Computer Science and Engineering, Guilin University of Technology, Guilin 541006, China.
- Department of Neurology, the 924th Hospital of the Joint Logisties Support Force of the Chinese People's Liberation Army, Guilin 541000, China.
- Hospital of Guangxi Normal University, Guilin 541004, China.
- Guangxi Key Laboratory of Embedded Technology and Intelligent System, Guilin University of Technology, Guilin 541006, China.
Abstract
Computed tomography (CT) scan is a widely used, non-invasive, sensor-based imaging technique that provides critical intracranial information for rapid stroke assessment. Accurate segmentation of intracerebral hemorrhage (ICH) in sensor-derived CT images is vital for clinical decision-making. Effective intelligent analysis of CT images is key to achieving reliable computer-aided diagnosis. However, existing deep learning methods struggle with complex ICH lesions characterized by blurred boundaries, irregular shapes, and large-scale variations. To address these challenges, this paper proposes TransAMGNet, a hybrid U-shaped network with Transformer integration for ICH CT image segmentation. The network is built on a residual U-Net backbone and introduces a Transformer encoder to strengthen global context modeling, thereby improving the representation of complex lesion morphology. Specifically, in the encoding stage, we design an Adaptive Dual-branch Channel Attention Module (ADCAM), which jointly models global and local channel information to enhance the model's sensitivity to important feature responses. In the skip-connection pathway, we introduce a Multi-scale Feature Enhancement Module (MFEM), which preserves high-resolution spatial details while supplementing multi-scale contextual information to improve shallow-deep feature fusion. During decoding, a Gate-enhanced Dynamic Upsampling Module (GDUM) is constructed to improve the recovery of lesion boundaries and fine-grained structures through the synergy of gated recalibration and content-aware upsampling. The proposed method is systematically evaluated through comparative experiments and ablation studies. Experimental results show that TransAMGNet outperforms competing methods across multiple evaluation metrics, achieving Dice, Recall, IoU, Precision, and HD95 values of 90.47 ± 0.58%, 87.83 ± 3.71%, 81.26 ± 0.78%, 91.13 ± 0.95%, and 32.94 ± 1.1, respectively. The ablation studies further verify the effectiveness of each module. These results demonstrate that TransAMGNet can effectively improve segmentation performance for complex ICH lesions.