ENSEMBLE OF IMPROVED EFFICIENTNET MODEL FOR MELANOMA DETECTION IN DERMOSCOPIC IMAGES

Melanoma has increased in prevalence over the last three decades, and early detection is critical for lowering the mortality rate linked to its type of skin cancer. Thus, access to an automated, dependable system capable of detecting the existence of melanoma using a dermatoscopic image of lesions can be an essential instrument in the field of clinical diagnostics. Among the state-of-the-art technologies for automated or computerized clinical diagnostics, deep learning based on Convolutional Neural Networks should given attention, which have been utilized to create classification and detection systems for various disorders. The method suggested in this article utilizes an imaging stage to generate the features of the dermatoscopic picture using the EfficientNet-based Convolutional Neural Network, followed by an attention mechanism to reweight the features. Following that, we utilize the dense layer to process metadata and build a fully connected layer that categorizes lesions as "benign" or "malignant". We have trained, validated and tested the suggested model using the database related to the 2020 International Symposium on Biomedical Imaging challenge. The ensemble model’s ROC-AUC is 0.9628, while its recall, precision, and F1-score are 0.8639, 0.8746, and 0.8692, respectively. Thus, it shows that the model can discriminate accurately between benign and malignant lesions without bias toward any clas