Implementation of a Pattern Classifier on Thermograms from Plantar Region

This study aims to implement a pattern classification algorithm for plantar thermograms, focusing on identifying altered temperature zones in the feet of diabetic patients. Utilizing a database of 334 thermograms, various classification algorithms including Support Vector Machine (SVM), Logistic Regression, Artificial Neural Network (ANN), Random Forest, and K Nearest Neighbors (KNN) were evaluated. Features extracted from the literature, such as number of pixels, maximum entropy, variance, mean value, correlation, contrast, energy, and homogeneity, were utilized for training and evaluation. The classification task involved assigning thermograms to 5 classes based on the thermal change index (TCI). Performance evaluation was conducted using an information theory metric approach based on mutual information, measuring the alignment between predicted and true classes. The neural network achieved the highest mutual information score of 2.69 out of 5, indicating that approximately 53.8% of the information obtained from model predictions aligned with the true classes. Additionally, a database was established in the Footlab BASPI laboratory, comprising 20 thermograms from the plantar region of 10 control subjects. A novel protocol, incorporating additional elements to the STANDUP base protocol, was proposed. Finally, classification using the ANN on data acquired from the Footlab - BASPI database yielded satisfactory results, successfully distinguishing between 1.7 classes, representing an 85% success rate in classifying thermograms.