Article Details

Abstract

Aiming at the problems of missed detection of critical structures, coarse segmentation boundaries, and parameter redundancy caused by variable vertebral morphology, complex background structures, and low tissue contrast in lumbar MRI, this paper proposes YOLOMACR-Net, a lightweight lumbar segmentation model integrating multi-scale feature enhancement and a cascade architecture. Firstly, a Multiscale Asymmetric Cavity Residual (MACR) module is designed within the YOLOv5n framework, utilizing asymmetric convolution to adapt to vertebral geometric features and expand the receptive field to address missed detections in single-stage detectors. Secondly, it constructs a “localization-segmentation” cascade architecture, uses the localization results to filter background noise, and guides U-Net for fine-grained segmentation. Experimental results on public datasets show that YOLOMACR-Net achieves a Structure Capture Rate (SCR) of 100%, with mIoU, Dice coefficient, and HD95 reaching 88.17%, 93.71%, and 3.37 mm, respectively, while the parameter count is only 1.65M. The results demonstrate that the model effectively integrates multi-scale information and significantly improves segmentation accuracy in complex scenes while maintaining a lightweight design.

Keywords

medical image segmentation; Deep Learning; YOLO; MACR; U-Net

References

1. [1] CUI Y Y，ZHU J，DUAN Z L，et al. Artificial Intelligence in Spinal Imaging: Current Status and Future Directions [J/OL].International Journal of Environmental Research and Public Health，2022，19（18）：11708[2025-09-06].https://doi.org/10.3390/ijerph191811708.
2. [2] RAYED M E，ISLAM S M S，NIHA S I，et al. Deep learning for Medical Image Segmentation: State-of-the-art Advancements and Challenges [J/OL].Informatics in Medicine Unlocked，2024，47：101504[2025-09-13].https://doi.org/10.1016/j.imu.2024.101504.
3. [3] ISENSEE F，JAEGER P F，KOHL S A A，et al. nnU-Net: A Self-configuring Method for Deep Learning-based Biomedical Image Segmentation [J].Nature Methods，2021，18（2）：203-211.
4. [4] CHEN J N，LU Y Y，YU Q H，et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation [J/OL].arXiv:2102.04306 [cs.CV].（2021-02-08）.https://arxiv.org/abs/2102.04306.
5. [5] CAO H，WANG Y Y，CHEN J，et al. Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation [J/OL].arXiv:2105.05537 [eess.IV].（2021-05-12）.https://arxiv.org/abs/2105.05537.
6. [6] VALANARASU J M J，PATEL V M. UNeXt: MLP-based Rapid Medical Image Segmentation Network [J/OL].arXiv:2203.04967 [eess.IV].（2022-03-09）.https://arxiv.org/abs/2203.04967.
7. [7] RUAN J C，XIANG S C，XIE M Y，et al. MALUNet: A Multi-Attention Light-weight UNet for Skin Lesion Segmentation [C]//2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM).Las Vegas：IEEE，2022：1150-1156.
8. [8] RUAN J C，XIE M Y，GAO J S，et al. EGE-UNet: An Efficient Group Enhanced UNet for Skin Lesion Segmentation [J/OL].arXiv:2307.08473 [eess.IV].（2023-07-17）.https://arxiv.org/abs/2307.08473.
9. [9] WU H S，CHEN S H，CHEN G L，et al. FAT-Net: Feature Adaptive Transformers for Automated Skin Lesion Segmentation [J/OL].Medical Image Analysis，2022，76：102327[2025-09-11].https://doi.org/10.1016/j.media.2021.102327.
10. [10] KAZEROUNI A，AGHDAM E K，HEIDARI M，et al. Diffusion Models in Medical Imaging: A Comprehensive Survey [J/OL].Medical Image Analysis，2023，88：102846[2025-09-11]https://doi.org/10.1016/j.media.2023.102846.
11. [11] GUO M-H，LU C-Z，HOU Q B，et al. SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation [J/OL].arXiv:2209.08575 [cs.CV].（2022-09-18）.https://arxiv.org/abs/2209.08575.
12. [12] LIU Z，MAO H Z，WU C-Y，et al. A ConvNet for the 2020s [C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).New Orleans：IEEE，2022：11966-11976.
13. [13] JOCHER G，CHAURASIA A，STOKEN A，et al. Ultralytics/yolov5:v7.0-YOLOv5 SOTA Realtime Instance Segmentation [EB/OL].（2022-11-22）[2025-09-12].https://doi.org/10.5281/zenodo.7347926.
14. [14] PANG S M，PANG C L，ZHAO L，et al. SpineParseNet: Spine Parsing for Volumetric MR Image by a Two-stage Segmentation Framework with Semantic Image Representation [J].IEEE Transactions on Medical Imaging，2021，40（1）：262-273.

About the authors

He Zhiyuan (1997-), male, Han ethnicity, from Fuzhou, Jiangxi Province, is currently pursuing a master's degree. His research focuses on deep learning based lumbar spine image segmentation; Corresponding author: Wang Canhua (1978-), male, Han ethnicity, born in Jiujiang, Jiangxi Province, associate professor, expert in medical device evaluation and consulting in Jiangxi Province, director of the Traditional Chinese Medicine Health Big Data Branch of the China Association of Traditional Chinese Medicine Information, doctoral degree, research direction: health big data and artificial intelligence, biomedical engineering, machine learning applications in medical instruments.