Heuristic Segmentation Assisted Deep-Spatial Feature Learning Model for Leprosy Detection

The high pace of rising leprosy cases and resulting socio-stigma has alarmed academia-industries to achieve non-invasive, touchless, and vision-based computer aided diagnosis (CAD) solution for leprosy detection and classification. Unlike classical CAD systems, leprosy detection and classification has remained a least explored research area. Invasive and chemical assay-based leprosy detection approaches are often found time-consuming, complex, and limited to cope-up with real-world’s demand. Though, in the last few years every effort has been made towards vision-based leprosy detection and classification; however, limited data, inferior feature, high-annotation demands, and importantly low accuracy limit their suitability. Considering these facts as motivation, in this paper a highly robust Heuristic Driven Segmentation assisted hybrid deep-spatio-textural feature learning for leprosy detection and classification is developed. It emphasizes on both region of interest segmentation and its optimization, as well as feature improvement to achieve higher accuracy and reliability. Here, firefly heuristic driven Fuzzy C-Means clustering (FFCM) was developed to perform ROI specific segmentation, which was followed by a hybrid deep spatio-textural feature extraction process. Noticeably, FFCM at one hand enabled automatic and accurate ROI-segmentation. On the contrary, the use of hybrid features including descriptive spatio-textural textural statistics (i.e., Gray-level co-occurrence metrics) and 4096-dimensional AlexNet features provided an intrinsic feature rich vector to perform accurate leprosy classification. The ROI-specific hybrid features (i.e., GLCM and AlexNet features) were processed for two-class classification using ensemble random forest classifier that labelled each input skin lesion as the normal image or the leprosy detected. The proposed leprosy detection method exhibited superior performance (accuracy = 96.6%, precision = 99.7%, recall = 95.8%, F-score = 0.9771) over other state-of-art methods like GLCM (accuracy = 88.51%) and Convolutional Neural Network based methods (91%). It affirms its suitability of the proposed CAD model towards automated and touchless real-time leprosy diagnosis.