Related Experiment Video
Updated: Sep 3, 2026
Multicolor 3D Printing of Complex Intracranial Tumors in Neurosurgery
Published on: January 11, 2020
Beyond 2D slices: TD-Mamba for 3D CT segmentation of head and neck space-occupying lesions
Thao Thi Phuong Dao1, Tan-Cong Nguyen2, Minh-Khoi Pham3
1University of Science, Ho Chi Minh City, Viet Nam; Thong Nhat Hospital, Ho Chi Minh City, Viet Nam; Vietnam National University, Ho Chi Minh City, Viet Nam.
Background And Objective:
Space-occupying lesions of the head and neck can obstruct the aerodigestive tract and involve critical neurovascular structures, potentially resulting in life-threatening complications such as sepsis or acute airway compromise. Despite their clinical importance, publicly available medical imaging benchmarks primarily focus on malignant tumors and largely overlook other clinically relevant lesion types.
Methods:
To address this limitation, we present 3D-HNSeg, a curated dataset of 204 head and neck CT scans with expert manually annotated voxel-wise 3D ground-truth of three lesion categories, including tumors, cysts, and abscesses. To address the substantial anatomical variability and heterogeneous appearance of these lesions, we propose TD-Mamba, a synergistic and spatially adaptive segmentation architecture. This model integrates Tri-oriented Dilated Mamba blocks for efficient multi-scale 3D contextual representation and Soft Signal-Adaptive Memory modules for token-wise feature refinement via adaptive memory gating.
Results:
Experiments on the 3D-HNSeg dataset demonstrate that TD-Mamba outperforms baseline methods, achieving a Dice Similarity Coefficient (DSC), mean Intersection over Union (mIoU), and 95th percentile Hausdorff Distance (HD95) of 36.31%, 30.10%, and 21.59 mm, respectively. Ablation studies further validate the contribution of each architectural component. Although TD-Mamba improves over the evaluated baselines, its absolute segmentation performance remains substantially lower than inter-observer agreement, particularly for abscesses. Therefore, the model should be interpreted as a research baseline rather than an autonomous clinical solution.
Conclusion:
These results demonstrate the effectiveness of TD-Mamba as a strong baseline on the proposed 3D-HNSeg benchmark, supporting future research on automated head and neck lesion segmentation and related computer-assisted clinical applications. The dataset and source code are available at https://github.com/drthaodao3101/3D-HNSeg.
