Skin cancer in particular melanoma is increasingly representing one of the biggest cause of death and real time detection through precise segmentation is important to enhance survival rate. This work presents a new enhanced SegNet architecture for skin lesion segmentation aimed at overcoming difficulties such as image variations, noise and overfitting. The deeper convolutional layers, multi-rate dropout regularization and multi-metric evaluation (including IoU and Dice Coefficient) are used to further enhance segmentation precision as well as model generalization. The model is evaluated on the PH2_resized dataset and compared with state-of-the-art methods such as U-Net and ResNet50; it achieves an IoU of 97.02% and an accuracy of 96.83%. The originality of the method lies in integration of regularization techniques, advanced feature extraction and complete evaluation metrics to improve robustness and performance for the segmentation. Such results indicate that we have developed an effective model of precise segmentation on skin lesions, which has promising potential for the applicability to automatic melanoma detection and clinical assessment.
Keywords: skin lesion segmentation, melanoma, deep learning, SegNet architecture, dropout regularization; performance metrics
How to cite this article: Voggu S, Siddiqui S, Deep Learning–Driven Melanoma Detection via Regularized SegNet Skin Lesion Segmentation. Int J Drug Deliv Technol. 2026;16(2s): 160-169; DOI: 10.25258/ijddt.16.160-169