Performance Evaluation of Handcrafted Feature Descriptors for Skin Disease Classification using Decision Tree Classifier

Effective dermatologic interventions in addition to clinical decisions entail a timely accurate identification of dermatological conditions. This study involves an analysis of the six manually designed image feature extracting algorithms, i.e. Auto Color Correlogram Filter (ACCF), Binary Patterns Pyramid Filter (BPPF), Fuzzy Color and Texture Histogram Filter (FCTHF), Fuzzy Opponent Histogram Filter (FOHF), Gabor Filter, and JPEG Coefficient Filter (JPEG CF), in collaboration with the Decision Tree classifier in the context of automating dermatological diagnosis. The analysis involved the utilization of images of the Skin Diseases Image Dataset of the Kaggle which comprises of over thirty-eight thousand pictures which span in ten main skin diseases categories. After every feature descriptor was used, the resulting processed image set was classified using the implementation of Decision Tree algorithm via WEKA 3.9.5 environment. Standard evaluation methods were applied in estimating the evaluation measures that comprised of Accuracy and also Precision, Recall, Receiver Operating Characteristic (ROC) and Precision-Recall Curve (PRC) besides Computation Time. The currently conducted evaluation indicated that the overall estimate in performance of JPEG CF in combination with a Decision Tree approach was 97.05 percent accuracy along with precision and recalls of 0.97 and ROC/PRC magnitude of 0.98. FOHF exhibited both a mixture of 95.55 percent rate of accuracy and 0.03 seconds of fastest classification. The proposed models hold the potential of being applied in the platforms of clinical decision support.