Hybrid Models for Medical Image Analysis: CNN-SVM in Diabetic Foot Ulcer Detection

Diabetic foot ulcer (DFU) is one of the most critical consequences of diabetes characterized by high mortality and potential risks for lower limb amputations. It is important if the disease is to be managed effectively that there is early and accurate detection however the current diagnostic technologies are effect inefficient and imprecise. This study proposes a new CNN-SVM model to mitigate these challenges owing to the extraction of pertinent features from the images using the CNN layers and classification accomplished by the SVMs. This proposed integrated CNN-SVM model yielded impressive performance by delivering an accuracy of 96.7 % which is higher than purely CNN based models (91.3%) and methods based on hand-designed feature representation (82.4%).Thus, the model is 94.8% accurate at detecting true positives and shows only 2.5% false-positive rates, which make it suitable for clinical use. The hybrid model also was performing weal with an accuracy and F1-score of 95.6%. Furthermore, the macro-average area under the curve for ROC was 98.2%; the area under the curve per DFU Class – Normal: 98.5%; Mild: 97.8%; and Severe: 98.3% – demonstrated a strong discriminator across varied severities of DFU. The model was trained using transfer learning on the DFUC2021 dataset, getting 97.0% of training accuracy and 96.7% of validation accuracy with little overfitting as evaluated by the losses of 0.08 and 0.10, consecutively. External validation on the IEEE DataPort Diabetic Foot dataset also supported IEEE DataPort Diabetic Foot dataset accuracy and its scalability. The results of this study show the good performance of the proposed CNN-SVM approach; this indicates that, the method is efficient, interpretable and could be extended to a large scale in regards to the automation of DFU detection, therefore contributing greatly towards enhancing the quality of life of patients with DFU and reducing health costs.