Hernández-Corvo method for footprints classification using artificial intelligence

Abstract

The use of artificial intelligence (AI) in data analysis and image processing has gained recognition for its high validity and reliability. The Hernández-Corvo Index method is widely used in podiatric diagnostics and related fields to classify plantar footprints. Therefore, the objective of this study was to classify plantar footprints using the Hernández-Corvo method by applying convolutional neural networks (CNNs) as a measurement tool. An observational design with a descriptive cross-sectional approach was used. A total of 275 subjects without mobility problems participated in the study. Prior to recording the plantar footprint, participants rested for 15 minutes to minimize foot vascularization. Footprints were classified as normal, cavus, or planus, and photographs were taken to generate a dataset of 551 images for CNN training. Different convolutional architectures were modeled, and techniques such as data augmentation and regularization were applied. Feature extraction was performed using transfer learning with the pre-trained VGG16 network. Several CNN models were trained using different techniques. The basic architecture achieved a validation accuracy of 53%. Incorporating L2 regularization and dropout in the central layers improved model performance in terms of overfitting control while maintaining validation accuracy. By unfreezing the last two layers of the CNN and using the VGG16 model, a validation accuracy of 91% was achieved. However, significant errors were still observed the classification of cavus and normal foot types. The feasibility of training a pre-trained network with limited computational resources has been demonstrated. The use of the smaller VGG16 network facilitated training without starting from scratch, highlighting the effectiveness of transfer learning in improving object detection results. Due to computational limitations, VGG16 facilitated image processing for each foot type, requiring only training of the final model layer. The use of CNNs to classify plantar footprints based on the Hernández-Corvo method in apparently healthy individuals has shown high efficacy.