Automated knee MRI segmentation with KneeSeg-U for cartilage and bone extraction using an unsupervised deep learning framework.

The segmentation of knee cartilage and bone in magnetic resonance images with supervised learning methods needs manual annotations from experts, although it demands high amounts of labeled data. KneeSeg-U presents itself as a deep learning unsupervised framework that extracts cartilage segments and bone structures from knee MR imaging automatically with unmarked scan data. KneeSeg-U provides automated full-system segmentation with precise results through unaided clinical efficacy and by negating the requirement for expert annotations. The creation of our extensive knee MRI dataset involved obtaining various sequences from both open-access repositories and clinical points of origin. During training, we applied domain-specific enhancement approaches along with contrastive learning algorithms to build better generalization abilities. The U-Net architecture applied in KneeSeg-U remains trainable through adversarial methods that incorporate self-supervised learning features with consistency restrictions for achieving precise segmentations. Experimental tests demonstrate KneeSeg-U produces segmentations that match supervised methods through a cartilage Dice similarity measure of 0.87, even as bone Dice similarity reaches 0.91 above typical unsupervised segmentation methodology. The system shows generalization functionality for different MRI protocols and multiple anatomical patterns. The KneeSeg-U framework represents an effective method to automate MRI knee segmentation without needing annotated references for building adaptable methods in clinical research about musculoskeletal disease diagnosis.