Automated Skin Lesion Diagnosis and Classification using K-mean, LAB-Color-Space Segmentation and Deep Learning

Effective and precise diagnosis tools are required due to the widespread occurrence of skin disorders globally; in addition, it takes a long time for analysis and classification. The diagnosis of skin diseases is complicated by the diversity of types, which is impacted by environmental, regional, and hereditary variables. This paper introduces a computerized diagnostic method that utilizes the pre-processed and segmented dataset and is applied to the latest version of the pre-trained MobileNet-v2 model. The segmentation process is based on the K-mean and LAB color space algorithm. The system underwent training and evaluation using the HAM10000 dataset, which consists of 10,015 pictures. The assessment of four pre-trained deep learning models, namely EfficientNet-b0, GoogLeNet, ShuffleNet, and MobileNet-v2, was conducted using metrics such as accuracy, precision, Jaccard index, sensitivity, specificity, and F1-score. MobileNet-v2 demonstrated exceptional performance with an accuracy of 91.43%, precision of 85%, Jaccard index of 81%, sensitivity of 83%, specificity of 93%, and F1-score of 82%. This approach assists in promptly identifying and categorizing skin disorders, enabling immediate and suitable medical intervention. Furthermore, the system can be seamlessly incorporated into apps to facilitate automated diagnosis by experts, especially for mobile healthcare systems.