Dermascopic and non-Dermascopic image classification using deep learning

Skin cancer, especially melanoma, is a serious threat to global health, with millions of new cases and considerable mortality each year. Early detection of cancerous lesions significantly improves patient outcomes, so the need for accurate diagnostic tools is evident. This work introduces a deep learning framework designed to classify dermoscopic images into malignant or benign categories. Our approach uses the architectures of MobileNetV2 and VGG-19 to obtain a classification accuracy of 94.06% on the ISIC dataset containing more than 10,000 dermoscopic images. For greater reliability, we apply intensive preprocessing techniques, such as data augmentation and balancing, for class imbalance mitigation. Apart from this we also worked on the ISIC 2024 3D TBP dataset, which is a dataset of non-dermascopic images that resemble close-up smartphone-like images, unlike the dermascopic dataset. We implemented an EfficientNetB0 model on this dataset to predict benign or malignant lesions. This proposed solution I adaptive and aims to work for both dermascopic as well as non-dermascopic images, thus providing a broader solution and includes mechanisms for providing feedback to enable continuous improvement. This project has, therefore, demonstrated high efficacy in transforming dermatological diagnostics and reducing cancer-related mortality rates, possibilities of deep learning for such applications.