Download PDFOpen PDF in browserQuantifying Segmentation Uncertainty to Evaluate the Quality of ML Generated Deltoid Masks in Shoulder Arthroplasty Patients4 pages•Published: January 5, 2026AbstractDespite the growing development of image-based machine-learning models, their integration into clinical practice remains limited. A significant barrier to adoption is the reliability of these models' predictions. This study demonstrates the use of uncertainty analysis to evaluate output of a CT-based model trained to segment deltoid muscles in shoulder arthroplasty patients. By quantifying uncertainty through metrics such as entropy, mutual information, and variance, we created 46 distinct image-level uncertainty scores for 108 good-quality and 100 low-quality segmentation outputs. In addition, these uncertainty scores were used to train a Gaussian Naïve Bayes model to identify low-quality cases, and the results were compared with those from single-metric thresholding. The results show that boundary 75 percentile entropy is the most predictive single uncertainly parameters (accuracy: 68%, recall: 68%, precision: 67%) while the trained model outperformed all single predictive metrics (accuracy: 78%, %, recall: 76%, precision: 78%). Our study indicates a uses case of utilizing uncertainty analysis to identify segmentation outputs that may require further manual correction, which will increase the trust, and potentially help for clinical adoption of ML segmentation models.Keyphrases: computer vision, shoulder arthroplasty, uncertainty analysis In: Joshua William Giles and Aziliz Guezou-Philippe (editors). Proceedings of The 25th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, vol 8, pages 139-142.
|
