AI; R-CNNAnalyse assistée par l'IA des échographies du genou pour la prédiction du résultat dans les affections dégénératives du genou

High-resolution musculoskeletal ultrasonography has become a first-line imaging modality because it enables real-time visualization and dynamic assessment with high accessibility and low cost. Nevertheless, ultrasound remains highly operator-dependent, resulting in variability in image acquisition and interpretation, which limits standardization and widespread implementation, particularly for complex joints such as the knee. Building on our established expertise in computational ultrasound and deep-learning-assisted dynamic shoulder analysis, including patented artificial intelligence (AI)-derived quantitative biomarkers, this three-year project aims to develop an AI platform for advanced knee ultrasound analysis and to construct a predictive model for clinical outcomes following ultrasound-guided injections in degenerative knee disorders. In the first year, we will establish a normative AI foundation model for knee ultrasonography by developing automated localization and multi-structure segmentation of major anatomical components, including tendons, ligaments, cartilage, fat pads, and peripheral nerves. Standardized acquisition protocols will be implemented to ensure consistent image quality. A Faster Region-Based Convolutional Neural Network (Faster R-CNN) framework incorporating ResNet50, a Feature Pyramid Network, and a Region Proposal Network will be used to detect key bony landmarks, followed by a multi-structure segmentation engine and quantitative feature extraction modules (e.g., thickness, surface regularity, and tissue heterogeneity). Segmentation performance will be evaluated using Intersection-over-Union and Dice coefficients, while measurement reliability will be assessed using intraclass correlation coefficients, standard error of measurement, minimal detectable change, and Bland-Altman analyses. In the second year, the platform will be expanded to differentiate pathological patterns in knee tendons, ligaments, cartilage, and fat pads. Expert clinicians will label each segmented structure as normal or abnormal and further annotate clinically relevant subtypes, such as tendinopathy, calcification, partial or full-thickness tears, synovial hypertrophy or effusion, cartilage wear or exposure, and meniscal degeneration or tear. Supervised learning models will be trained for classification and evaluated using accuracy, precision, recall (sensitivity), and F1-score. In the third year, we will develop an outcome prediction model for ultrasound-guided injections by integrating retrospective and prospective real-world data from approximately 150 patients receiving common injection therapies, including intra-articular hyaluronic acid, dextrose prolotherapy or platelet-rich plasma, and peripheral nerve-targeted interventions. Treatment success will be defined using validated patient-reported outcome measures, including the Knee Injury and Osteoarthritis Outcome Score and the Patient Acceptable Symptom State, incorporating minimal clinically important difference thresholds. Feature selection methods and cross-validation will be applied to mitigate overfitting. Model performance will be assessed using area under the receiver operating characteristic curve, sensitivity, specificity, accuracy, F1-score, and calibration metrics, with Shapley Additive exPlanations employed to enhance interpretability. External validation will be performed if additional datasets become available. This project is expected to deliver the first systematic AI-based normative atlas for knee ultrasonography and an interpretable outcome prediction framework, improving diagnostic consistency, reducing operator dependency, and enabling personalized, evidence-informed injection strategies.