Enhanced Skin Disease Detection Using Deep Learning with Fully-Convolutional Residual Networks and Lesion Index Calculation Unit

Detection of skin diseases is a very important branch of medical image analysis and, hence, early diagnosis is vital for the patient's recovery. This paper proposes an integrated multi-stage pipeline of deep learning for automated skin lesion analysis, which includes lesion segmentation, coarse classification, and refined classification using a Lesion Index Calculation Unit (LICU). The segmentation stage precisely segments the lesion, which has a high boundary accuracy of 95%, thus, data that are restricted to real sections could be analyzed. The coarse classification stage is a preliminary step of disease probability, but it has a high tendency for false positives. The LICU refinement stage not only removes these overdiagnoses but also increases the sensitivity by 0.92 and improves the accuracy to 92^%, thus, precision in lesion detection is successfully achieved. Quantitative metrics such as sensitivity, specificity, and the time it takes to process the data, verify both the robustness and the speed of the model. The qualitative analysis also proves that the refined classification stage is more accurate and therefore goes beyond lesion boundaries, which is essential for clinical diagnostics. This multi-step technique not only sharpens the doctor's tools for recognizing the disease but also offers a model that can be easily interpreted thus facilitating doctors in the analysis of complex cases such as skin conditions. In addition, the proposed pipeline is useful in the clinical sector by enabling precise and faster skin disease diagnosis, according to our experiments.