Detection of osteochondral lesion of talus in ankle magnetic resonance images with GradCAM-based hybrid CNN model

Abstract Background Treatment of osteochondral lesion of talus (OLT), one of the crucial pathologies that can cause pain in the ankle, is guided by the age of onset, severity, and stage of the symptoms. For these reasons, early screening and early intervention for OLT become important. Magnetic resonance imaging (MRI) is used for further evaluation. However, depending on the clinician’s experience, the diagnostic accuracy of the same images varies between physicians. In this study, we tried to determine the presence or absence of OLT using an AI-based hybrid model. Methods This study applied Gradient-Weighted Class Activation Mapping (GradCAM), an image visualization technique, to OLT images. Features were extracted and combined from the original and GradCAM-applied images. Then, the most valuable features from this high-dimensional feature map were selected using the Neighborhood Component Analysis (NCA) dimension reduction method. In the last stage, the feature map with selected features was classified in the K-Nearest Neighbors (KNN) classifier. Results To compare the performance of our proposed model, feature extraction was performed with six pre-trained models accepted in the literature. These features were classified into six different classifiers. As a result, the proposed model achieved the highest success rate of 98.60%. The proposed hybrid model for detecting talus osteochondral lesions in ankle magnetic resonance images has obtained successful results. Thanks to this computer-aided system, experts’ workload will be reduced, and this system can be used in places without experts.