Early-Stage Melanoma Cancer Diagnosis Framework for Imbalanced Data From Dermoscopic Images

Skin problems are a serious condition that affects people all over the world. Prolonged exposure to ultraviolet rays' damages melanocyte cells, leading to the uncontrolled proliferation of melanoma, a form of skin cancer. However, the dearth of qualified expertise increases the processing time and cost of diagnosis. Early detection of melanoma in dermoscopy images significantly enhances its chance of survival. Pathologists benefit substantially from the precise and efficient melanoma cancer diagnosis using automated methods. Nevertheless, the diagnosis of melanoma has consistently been a challenging procedure due to the imbalance images and limited data. Our objective was to employ a novel deep method to diagnose melanoma from dermoscopic images automatically. The research has proposed a novel framework for detecting skin malignancies. The proposed plan, which includes CNN, DenseNet, a batch normalization layer, maxpooling, and a ReLU layer activation function, solves the overfitting problem well. Furthermore, we used a large number of samples for testing and effectively employed data augmentation to prevent any issues related to class imbalance. The Adam optimizer is the most efficient deep learning optimizer for addressing challenges associated with large datasets, such as lengthy processing times. This is due to its specifically designed algorithm. Experiments ensure that the proposed framework achieved 95.70% micro average accuracy on the ISIC-2019 dataset and 93.24% accuracy on the HAM-10000 dataset. Comprehensive evaluation and analysis were used to evaluate our framework's performance. The results show that the proposed approach performs better with cross-validation by 94.8% accuracy than the most sophisticated deep learning-based technique. During studies, medical professionals will employ the proposed model to identify skin cancer in its early stages.