A Review On: Deep Learning Model For Skin Lesion Classification Using Imbalance Dataset

Deep convolutional neural networks (DCNNs) have undergone extensive research for the identification of skin diseases, and some of them have demonstrated diagnostic accuracy superior to that of dermatologists. However, small size and data imbalance are two issues that prevent large-scale implementation of DCNNs for skin-disease detection. A single-model approach has been proposed for skin lesion classification on small as well as imbalanced datasets. In addition, to verify the fact that moderate complexity DCNNs outperform large models, several DCNNs are trained on imbalance datasets. After this regularization Drop-Out and Drop-Block has been added to reduce overfitting, which is a significant issue in small datasets. The modified Rand Augmentation strategy is proposed to reduce overfitting and address sample underrepresentation. Further, in order to address the problems of variable sample size and classification complexity and to lessen the effect of anomalous samples on training, a distinctive multi-weighted new loss function as well as an end-to-end cumulative learning strategy(CLS) are presented. Our single DCNN approach, when combined with Modified Rand Augment, the MWNL, and the CLS, produced classification accuracy on dermoscopic image datasets that was comparable to or better than theirs. Additionally, given that it operated at a lower cost in terms of processing resources and inference time, this technique might be applied in the mobile devices for automated diagnostics of skin lesions and the malignancies.