Derma AI: A Two-Stage System for Skin Lesion Classification Using U-Net Segmentation and Multi-Modal Feature-Enhanced EfficientNetV2

Early and accurate diagnosis of skin lesions, particularly malignant melanoma, is critical for improving patient survival rates. Traditional visual inspection is subjective, and standard computer-aided diagnosis (CAD) systems often lack robustness to noise or fail to capture subtle diagnostic features. This paper presents a novel, two-stage deep learning system for automated skin lesion analysis. First, a U-Net model performs precise semantic segmentation to isolate the lesion from surrounding skin, mitigating noise from artifacts. Following segmentation, a powerful classification model, EfficientNetV2, is used for identification. The key contribution of this work is the enhancement of the classifier’s input with a multi-modal feature set, including Micro DWT (Discrete Wavelet Transform) for texture analysis, edge detection metrics for boundary irregularity, and advanced RGB channel statistics for color variegation. This hybrid approach allows the model to capture subtle cues missed by standard end-to-end models. Experimental results on the ISIC 2019 dataset demonstrate the system’s efficacy, achieving a Dice Coefficient of 0.86 for segmentation and a classification accuracy of 92.5%, with a 0.94 recall for melanoma. This system, designed for deployment as a scalable API cluster, offers a robust and accurate tool for clinical integration.