3D-SCUMamba: An Abdominal Tumor Segmentation Model

Identification and segmentation of tumors from CT scans are essential for early detection and effective treatment but they remain challenging due to imaging artifacts and significant variability in tumor location, size, and morphology. Existing deep learning models typically adopt encoder-decoder architectures integrating convolutional layers with global dependency modeling to capture broader contextual information around tumors. Capturing global dependencies is critical, as local information alone often inadequately distinguishes tumors, especially given their diverse sizes, complex shapes, and interference from imaging artifacts. Recent advancements observe the emergence of the convolution-transformer architecture, which improves segmentation performance on the cost of substantial computational resources. In this article, we propose 3D-SCUMamba, an innovative architecture strategically integrating State Space Modeling-Based deep learning (Mamba) within the bottleneck of the encoder-decoder structure to overcome the limitations of existing segmentation networks. The proposed model efficiently models global dependencies while maintaining stable training dynamics and efficient inference. Additionally, we introduce a novel Spatio-Context (SC) module utilizing 3D convolutions without pooling to enhance feature representations and reduce information loss commonly associated with pooling operations. The SC module effectively prepares summarized features for robust global dependency processing by the Mamba component. 3D-SCUMamba explicitly addresses the prevalent limitations in current deep learning methods by prioritizing robust feature representation, training stability, and substantial accuracy improvements. Comprehensive evaluations conducted on three medical imaging datasets: MSD Pancreas Tumor, MSD Colon Tumor, and Synapse BTCV, demonstrate that computationally efficient 3D-SCUMamba consistently outperforms state-of-the-art methods with segmentation accuracy improvements of 2% to 7%, exhibiting stable training behavior and efficient inference suitable for clinical applications. The code and dataset are available at https://github.com/juwita-sj/3DSCU-Mamba