Cape Town - 2026 ISMRM-ISMRT Annual Meeting and Exhibition - Segmentation for Musculoskeletal MRI Applications

Segmentation for Musculoskeletal MRI Applications

Digital Poster

Analysis Methods

Monday, 11 May 2026

Digital Posters Row I

09:15 - 10:10

Session Number: 368-02

No CME/CE Credit

Similar Sessions

This digital poster session presents recent advances in segmentation methodologies for musculoskeletal (MSK) MRI, spanning joints, bones, cartilage, muscle, and multi-tissue analysis. Methodological themes include atlas-based and shape-model approaches, diffusion- and wavelet-driven augmentation, semi-supervised and transformer-based learning, LLM-enhanced multimodal networks, and interactive user-in-the-loop segmentation. These contributions emphasize robust, data-efficient, and clinically meaningful MSK segmentation, supporting quantitative biomarkers, reproducibility, and improved workflow integration.

Skill Level: Intermediate

		368-02-001. Automated Hip Joint Segmentation on 3D Dixon MRI Using a High-Resolution Atlas Eros Montin, Iman Khodarahmi, James Lee, Riccardo Lattanzi NYU Grossman School of Medicine, New York, United States of America Impact: A high-resolution Dixon MRI atlas enables propagation of detailed cartilage anatomy onto standard clinical scans, overcoming visibility limits. The framework supports reproducible radiomic and biomechanical analyses in femoroacetabular impingement and provides ground-truth data for future deep-learning segmentation.
		368-02-002. LLM-Enhanced Multi-modal Network for Tibiofemoral Joint Tissue Segmentation in Knee MRI Lu Wen, Junru Zhong, Weitian Chen The Chinese University of Hong Kong, Hong Kong, China Impact: We presented a multi-modal segmentation framework that integrates textual and visual features for tibiofemoral joint tissue segmentation which could enhance diagnostic accuracy and drive further research in joint tissue analysis.
		368-02-003. PIDRP: joint reconstruction of cerebrovascular territorial mapping and feeding arteries labeling location of rVE-ASL Yining He, Jianing Tang, Tianrui Zhao, Ziyi Huang, Sang Hun Chung, Lirong Yan Northwestern University, Chicago, United States of America Impact: A deep learning pipeline was developed for robust and objective estimation of vascular territory maps and corresponding feeding-artery locations of rVE-ASL with fewer encoding steps. Compared with conventional perception-based methods, it achieves superior performance, allowing faster rVE-ASL signal acquisition.
		368-02-004. Learning Diffeomorphic Augmentations via Variational Latent Modeling: A Study on 3D MRIs of the Knee Joint Tonmoy Hossain, Bruno Astuto, Ashok Vardhan Addala, Fei Tan, Keerthi Sravan Ravi, Miaomiao Zhang, Ravi Soni University of Virginia, Charlottesville, United States of America Impact: Learning deformation patterns directly from clinical knee MRIs enables anatomically accurate, tissue-specific augmentations that improve segmentation performance. Our framework outperforms baseline methods by up to $5.5\%$ Dice, especially in data-scarce scenarios, advancing AI robustness in medical imaging.
		368-02-005. A neural shape model trained on 4,789 computed tomography vertebrae improves clinical magnetic resonance shape reconstruction Kathryn Marusich, Garry Gold, Akshay Chaudhari, Anthony Gatti Stanford University, Stanford, United States of America Impact: Neural shape models reduce bone shape reconstruction error of vertebrae segmented from low-resolution 2D clinical MRI by 13-23% even when trained on CT data, demonstrating the utility of NSMs trained on other modalities when reconstructing anatomical shapes from clinical MRI.
		368-02-006. Automated Multi-Tissue Knee Segmentation from Low-Resolution Dynamic UTE MRI using 3D U-Net Aayush Nepal, Nicholas Brisson, Jürgen Reichenbach, Martin Krämer Jena University Hospital, Jena, Germany Impact: We present a purpose-built, multi-tissue segmentation pipeline for low-resolution dynamic 3D UTE knee MRI. It automates bone and soft-tissue segmentation across motion cycles, reducing annotation time and improving reproducibility, enabling kinematic and soft-tissue analyses in both clinical and research settings.
		368-02-007. Comprehensive Quantitative MRI Biomarker Platform for Knee Osteoarthritis: Technical Validation and Multi-Reader Reliability Greg Gilles, Dagoberto Robles, Nathaniel Christiansen, Karim Jayyusi, Ali Shaikh, Rongrong Tang, Mihra Taljanovic, Jeffrey Duryea, Edward Bedrick, C. Kent Kwoh University of Arizona, Tucson, United States of America Impact: Knee osteoarthritis is burdensome, pervasive and lacks approved treatments, but quantitative biomarkers, specifically effusion-synovitis and bone marrow lesions, may bridge pathogenesis to therapies. Our rapid, reliable and scalable segmentation will unlock these biomarkers’ significance and facilitate targeted intervention and prevention.
		368-02-008. Data-Efficient Carpal Bone Segmentation in Wrist MRI localizers: Semi-Supervised versus Transformer-Based Feature Learning Ashish Saxena, Gurunath Reddy Madhumani, Chitresh Bhushan, Dattesh Dayanand Shanbhag GE HealthCare, Bengaluru, India Impact: This work advances data-efficient segmentation for wrist MRI by leveraging few-shot learning and self-supervised features, enabling scalable annotation with minimal labels. It improves automation in musculoskeletal imaging workflows, with semi-supervised models outperforming feature-based approaches.
		368-02-009. Interactive Semi-Supervised Segmentation of Knee Cartilage and Menisci Using a Hybrid Prompted U-Net on 7T T2* MRI Eisa Hedayati, Abdul Wahed Kajabi, Karsten Knutsen, Collin Steinberger, Abhinav Lamba, Stefan Zbyn, Asif Abul Hassan, Hassan Ahad, Jutta Ellermann Center for Magnetic Resonance Research (CMRR), Minneapolis, United States of America Impact: Prompt-guided semi-supervised learning enables accurate cartilage and meniscus segmentation on 7T T2* MRI with minimal manual input, providing an efficient pathway toward scalable, annotation-assisted quantitative musculoskeletal imaging.
		368-02-010. Automated Segmentation of Thigh Muscles in Polyneuropathies Kaizhong Shi, Ying Wang, Zichun Zhong, Jesus Fajardo, Hasan Sawan, Bo Hu, Jun Li, Yongsheng Chen Wayne State University School of Medicine, Detroit, United States of America Impact: The 3D full-resolution nnU-Net enables precise, reproducible segmentation of individual thigh muscles even in severely denervated muscles. The automated tool is ready to use for neuromuscular MRI research and clinical translation.
		368-02-011. Deep Learning MRI Segmentation for Automated Quantification of Leg Tissue Volumes and Fluid Distribution in Lymphedema Zola Bzdek, Nick Tustison, Rachelle Crescenzi, Klaus Hagspiel, Sheau-Chiann Chen, Shannon Taylor, Vanessa Crain, Allison O'Brien Scott University of Virginia, Charlottesville, United States of America Impact: This work presents an automated MRI analysis framework for tissue-specific leg segmentation and image quantification of objective metrics for lymphedema assessment. The long-term goal is to establish standardized imaging methods for studying fat- and fluid-related physiology to monitor lymphatic diseases.
		368-02-012. Optimizing Cartilage Segmentation with Conditional Wavelet Diffusion-Driven Knee MRI Contrast Augmentation Ashok Vardhan Addala, Fei Tan, Bruno Astuto, Ravi Soni GE HealthCare, San Ramon, United States of America Impact: By applying conditional wavelet diffusion to 3D knee MRI data, we generate high-resolution synthetic low-fat-sat images from fat-sat images that can augment training datasets for cartilage segmentation models, improving their accuracy and robustness to variations in fat saturation efficiency.
		368-02-013. Quantifying User Satisfaction: A Weighted Metric Approach for Evaluating Deep Learning-Based MRI Segmentations Falko Ensle, Ilker Özgür Koska, Nina Derron, Ulf Bach, Cagan Koska, Philipp Maintz, Marta Porta-Vilaro, Jonas Kroschke, Philipp Gerber, Roman Guggenberger Diagnostic and Interventional Radiology, University Hospital Zurich, University Zurich, Zurich, Switzerland Impact: The Likert estimation model can serve as a quality control layer for MRI segmentation algorithms, filtering out insufficient segmentations and averting time-consuming manual reviews. The weighted combination model allows to assess user satisfaction, instead of relying solely on individual metrics.
		368-02-014. Leveraging Organ Co-Occurrence Loss to Enhance Multi-Organ Segmentation in MRI Qianqian Qi, Runsheng Chang, Yanan Wu, Xiaoyun Liang Neusoft Medical Systems Co. Ltd, Shenyang, China Impact: The proposed U-Net-based architectures for medical image segmentation could improve multi-organ segmentation, thereby significantly reducing the workload of clinicians and enhancing diagnostic efficiency.