Machine Learning for Automated Staging of Progressive Collapsing Foot Deformity Using Weightbearing CT

Research Type: Level 3 - Retrospective cohort study, Case-control study, Meta-analysis of Level 3 studies Introduction/Purpose: In Progressive Collapsing Foot Deformity (PCFD), the determinants are multifactorial, originating in bony and soft-tissue involvement. Classification of PCFD requires an initial staging into flexible (stage 1) or rigid (stage 2), primarily based on clinical examination. However, this evaluation is subjective and influenced by physician experience and patient-specific factors. An objective data-based approach could enhance staging accuracy. Automated reports of multiple measurements can be generated from Weightbearing CT (WBCT) images, while Machine learning (ML) offers a promising approach to analyze those complex datasets. The present study aims to analyze the staging accuracy of WBCT based ML in PCFD cases. We hypothesized that substantial agreement would be found between clinical and ML. Methods: This retrospective cross-sectional study analyzed 73 feet with available WBCT datasets, staged as flexible or rigid and classified by a senior foot and ankle orthopaedic surgeon. Reports containing 37 measurements for each foot were produced after automatic segmentation. Demographics were also recorded. The dataset was normalized for uniform scaling (means set to 0 and standard deviations to 1) to improve model training convergence speed. Principal Component Analysis (PCA) reduced dimensionality, retaining 17 components of which 9 measurements accounting for 95% cumulative variance. Multiple Machine learning models were applied, including Logistic Regression, Random Forest Classifier, Voting Classifier, and Support Vector Classifier (SVC) of Polynomial Kernel degree = 4. The dataset was split into 80/20 samples for training and validation. Performance of models was evaluated using F1-score, precision, recall, accuracy and Area Under the Curve (AUC). Primary outcome was selection of models with F1-score greater than 0.80. Results: Mean age was 59.2y±15.6, BMI 32.6Kg/m2±6.4 and hindfoot alignment was Foot Ankle Offset 10.6%±4.9. There were 39 stage 1, flexible feet and 34 stage 2, rigid. The dataset used had 73 sample rows and 17 feature columns after dimensional reduction using PCA. Six columns were categorical and 11 were continuous. The CatBoost model demonstrated the highest overall performance, achieving an F1-score of 0.84, precision of 0.89, recall of 0.80, 0.80 of accuracy, and 0.76 AUC. Other notable models included SVC which had the best the AUC (0.82) and recall (1.0), LightGBM, and NaiveBayes. Conclusion: Machine Learning models could accurately stage PCFD using 3D WBCT-derived automated multiple measurements reports. Our approach demonstrated high performance across models, consistent through the training set split and secondary analysis. This suggests that WBCT-derived 3D measurements contain key multidimensional data that can be harnessed for PCFD staging. Machine Learning could support initial clinical evaluation by providing additional objective data-driven validation. Limitations included a small validation set, a low feature to sample ratio, and reliance on clinical examination as the benchmark. Future research should explore larger datasets and seek multi-center validation.