Page 56 - AIH-2-1

Artificial Intelligence in Health Asymmetric U-Net for enhanced spinal MRI segmentation

Figure 5. Segmentation maps of different models

segmentation, providing a practical solution that balances Acknowledgments
performance with computational efficiency.
None.
4. Conclusion Funding
In this study, we introduce a novel asymmetric U-Net
architecture, termed J-Unet, designed for efficient None.
and accurate spine MRI image segmentation. J-Unet
distinguishes itself from conventional U-Net and its Conflict of interest
variants through its asymmetric encoder-decoder The authors declare that they have no competing interests.
structure, featuring a deeper upsampling path that
enhances the precise reconstruction of anatomical details. Author contributions
The incorporation of adjacent-scale skip connections Conceptualization: Longfei Zhou
and PRCs allows J-Unet to reduce the number of model
parameters while maintaining the flow of multi-scale Formal analysis: Xingyu Chen, Weihao Cheng, Zhanghao
contextual information. We rigorously evaluated J-Unet Qin, Tianao Shen, Pingyu Cao, Zebo Huang, Xiangyu
using a dataset comprising 215 spine MRI images. The Wu, Yiyao Zhang
results indicate that J-Unet significantly outperforms Investigation: Longfei Zhou, Xingyu Chen
other models, including U-Net, U-Net++, and U-Net+++, Methodology: Xingyu Chen, Weihao Cheng, Zhanghao
achieving at least a 0.24% improvement in both dice score Qin, Tianao Shen, Pingyu Cao, Zebo Huang, Xiangyu
and mIoU. Furthermore, J-Unet operates with substantially Wu, Yiyao Zhang
fewer parameters compared to U-Net+++ and Res-UNET, Writing–original draft: Xingyu Chen, Weihao Cheng,
demonstrating superior performance and efficiency. In Zhanghao Qin, Tianao Shen, Pingyu Cao, Zebo
conclusion, this study presents an innovative asymmetric Huang, Xiangyu Wu, Yiyao Zhang
U-Net architecture specifically tailored for spine MRI Writing–review & editing: Longfei Zhou
image segmentation. J-Unet’s unique design enables
precise localization and segmentation while optimizing Ethics approval and consent to participate
parameter efficiency, making it a highly accurate and
resource-effective solution. Our model offers a promising Not applicable.
advancement in automating spine segmentation in medical Consent for publication
image analysis, potentially enhancing diagnostic processes
and treatment planning. Not applicable.

Volume 2 Issue 1 (2025) 50 doi: 10.36922/aih.3889