Detection of flatfoot condition from X-ray images using image filtering and transfer learning approaches

Flatfoot is a deformity that occurs as a result of the loss and flattening of the arch of the foot as a result of the weakening of the ligament, tendon and muscle structures that keep the bone and joint tissues in the foot in a certain order and curve due to various factors. Various health problems affecting the musculoskeletal system can restrict the person's movements in work life and daily social life. In this sense, one of the most common musculoskeletal problems is flatfoot. In this study, automatic detection of flatfoot is proposed using image filtering methods and transfer learning approaches. In this study, X-Ray images are enhanced with average, Gaussian, median and dilatation filtering methods and resized for each transfer learning architecture. Then, the performance of the transfer learning approaches DarkNet-19, DenseNet-201, GoogLeNet, MobileNetV2 and ResNet-101 are compared. As a result of the experiments, DenseNet-201 architecture with dilation filtering showed the highest performance. The proposed model achieved a high performance with 98.02% accuracy, 98.10% f1 score, 98.47% sensitivity, 97.54% specificity, 97.73% precision, 96.04% Matthew Correlation Coefficient (MCC) and 96% Cohen's Kappa statistic.