Efficient LoRA Adaptation of SAM3 for Interactive Bone Segmentation in Visible Human Cryosection Images

• Philipi Schneider UFSC
• Carlos Eduardo Weiss UFSC
• Eloiza de Oliveira Rosa UFSC
• Vinícius Souza Capistrano UFSC
• Eliane Pozzebon UFSC

DOI: https://doi.org/10.5753/eniac.2026.30882

Resumo

This study evaluates Low-Rank Adaptation (LoRA) of SAM3 for interactive bone segmentation in Visible Human cryosection images. Adapters trained on Visible Korean bone data were externally evaluated on right-knee and right-foot volumes using idealized reference-derived point prompts and axial propagation. The best adapters increased volumetric Dice from 0.739 to 0.971 in the knee and from 0.411 to 0.917 in the foot while reducing severe failures and 3D fragmentation. These results indicate that parameter-efficient adaptation is effective under this idealized protocol, although performance varies with anatomy, training data, and adapter configuration.

Referências

Ackerman, M. J. (1995). Accessing the visible human project. D-Lib Magazine.

Ackerman, M. J. (1998). The Visible Human Project. Proceedings of the IEEE, 86(3):504–511.

Andreassen, T. E., Hume, D. R., Hamilton, L. D., Walker, K. E., Higinbotham, S. E., and Shelburne, K. B. (2023). Three dimensional lower extremity musculoskeletal geometry of the Visible Human female and male. Scientific Data, 10:34.

Carion, N. et al. (2025). SAM 3: Segment anything with concepts. arXiv:2511.16719.

Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., and Zagoruyko, S. (2020). End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213–229.

Clunie, D. et al. (2024). DICOM converted images for the NLM-Visible-Human-Project collection. Zenodo. DOI: 10.5281/zenodo.12690050.

Hu, E. J. et al. (2022). LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR).

Kirillov, A. et al. (2023). Segment Anything. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992–4003.

Mazurowski, M. A. et al. (2023). Segment Anything Model for medical image analysis: An experimental study. Medical Image Analysis, 89:102918.

Meta AI (2026). Segment Anything Model 3 official implementation. [link].

Park, J. S., Chung, M. S., Hwang, S. B., Lee, Y. S., Har, D.-H., and Park, H. S. (2005). Visible Korean Human: Improved serially sectioned images of the entire body. IEEE Transactions on Medical Imaging, 24(3):352–360.

Park, J. S., Chung, M. S., Hwang, S. B., Shin, B.-S., and Park, H. S. (2006). Visible Korean Human: Its techniques and applications. Clinical Anatomy, 19(3):216–224.

Ravi, N. et al. (2024). SAM 2: Segment anything in images and videos. arXiv:2408.00714.

Spitzer, V., Ackerman, M. J., Scherzinger, A. L., and Whitlock, D. (1996). The visible human male: A technical report. Journal of the American Medical Informatics Association, 3(2):118–130.

University of Denver Center for Orthopaedic Biomechanics (2022). 3D models of the Visible Human male and female lower extremities. Dataset. DOI: 10.56902/COB.vh.2022.0.

Wu, J. et al. (2023). Medical SAM Adapter: Adapting segment anything model for medical image segmentation. arXiv:2304.12620.