Enhanced Flatfoot Classification With Transformer Encoders and Bayesian Optimized SVM Parameters

Early detection of diseases allows for timely interventions that can reduce the risk of severe health complications; using Machine Learning (ML) tools to support medical specialists aids in early detection. Flatfoot is a musculoskeletal disorder that affects the biomechanical structure of the foot and the distribution of pressure while walking or running, often resulting in pain or injuries in the lower extremities. This work presents a methodology for classifying foot types using ML algorithms, requiring minimal information to be effective. Ten algorithms were trained on a dataset of 626 images and evaluated using a test set of 268 images. Hyper-parameter tuning of ten algorithms is achieved through Bayesian Optimization (BO), utilizing a descriptor based on the attention mechanism of a vision transformer. The results obtained by optimizing the Radial Basis Function Support Vector Machine (RBF-SVM) with BO are promising for flatfoot disease, achieving an accuracy greater than 95% in foot type classification. The methodology is also implemented in biomedical datasets to compare the results obtained with Flatfoot and define a methodology that provides good results for the type of flatfoot images analyzed.