Abstract

Early melanoma detection is critical to improving patient survival, yet accurate screening remains dependent on specialized clinical expertise. SPECTRA is a deep learning prototype for melanoma classification using transfer learning with the InceptionV3 convolutional neural network. Leveraging pretrained feature representations, the model was trained and validated on a dermoscopic image dataset to perform binary lesion classification (benign vs. malignant). Data augmentation, dropout regularization, and class balancing strategies were applied to enhance robustness and mitigate overfitting. Performance was evaluated using AUC-ROC analysis. Results signal strong classification performance, highlighting the effectiveness of automated classification for biomedical imaging tasks, especially in underserved environments. SPECTRA aims to contribute a scalable, efficient pipeline for computer-aided skin cancer screening in resource-limited healthcare environments.