Improving the classification performance of a deep learning algorithm for pes planus diagnosis using metaheuristic optimization techniques

Pes planus is a foot deformity that occurs when the arch of the foot collapses, losing its normal structure, and can lead to pain, posture problems, and walking difficulties over time. This study aims to develop a deep learning-based classification system for the automatic detection of pes planus from foot images and to optimize this system using metaheuristic algorithms. The dataset used consists of a total of 558 images, including 137 normal and 421 pes planus images. To reduce data imbalance, the 137 images belonging to the normal class were augmented using data augmentation methods and expanded to a number close to that of the pes planus class. In the first stage, a basic CNN model was created using a MobileNetV2-based transfer learning approach and trained with standard hyperparameters. The test accuracy of this initial model was obtained as 83.6%. In the next stage, the model's hyperparameters were optimized using three different metaheuristic algorithms: genetic algorithm (GA), grey wolf optimization (GWO), and particle swarm optimization (PSO). The model was retrained for each hyperparameter combination generated by each algorithm, and performance results were calculated. As a result of optimization experiments, the test accuracy of the CNN model obtained with GA reached 89.1%, the accuracy of the model optimized with GWO reached 85.9%, and the accuracy of the model optimized with PSO reached 89.8%. These findings show that addressing data imbalance and hyperparameter optimization significantly improved classification performance.