DermSegNet: smart IoT model for multi-class dermatological lesion diagnosis using adaptive segmentation and improved EfficientNetB3

Subjective visual examination by human dermatologists is associated with inter-observer variability and error. To address this problem, we present a method to accurately diagnose the 23 most common skin conditions, using an adaptive GrabCut approach with the EfficientNetB3 model, for accurate segmentation and classification, respectively. Using a custom loss function, this strategy is combined with data-level preprocessing employing algorithm-level approaches. The unbalanced Dermnet dataset, which includes 19,500 images of skin lesions representing the 23 most common skin conditions, was corrected by downsampling the major classes and enhancing the minor classes. The custom loss function and ADG significantly enhanced model accuracy by accurately segmenting regions of interest in the images, retaining the most relevant diagnostic information. The MSE, PSNR and Jacquard index support the best segmentation result with values 32.94, 70.23, 0.71 respectively. The results demonstrated very high accuracy, with top-5 and top-1 accuracy rates of 95% and 80.02%, respectively. The diagnoses of acne and nail fungi were exceptionally good, with precision rates exceeding 0.80 and 0.856, respectively. Our Dermnet-trained model is the most accurate of all state-of-the-art models published to date.