WELDONNet: Weighted Ensemble DenseNet121, MobileNetV2, and NASNetMobile to improve skin lesion diseases classification

Abstract

Abstract Skin cancer is classified as an aggressive type of cancer spreading quickly to other organs and tissues. Thus, inappropriate detection of the condition might lead to mortality. Melanoma is the most lethal variety, as it can extend to all human body parts. Early detection by dermatologists is, however, challenging in terms of consistency and timing. Hence, CAD systems have been utilized. Yet, the problems often encountered include class imbalance and noise. Therefore, the present authors propose a new method for detecting multi-class skin lesions by employing weighted ensemble convolutional neural networks (CNN). This study consisted of several steps, firstly, up-sampling and down-sampling of datasets were implemented to overcome the class imbalance problem, and image resizing was performed to scale down the image pixels in order to decrease noise. Secondly, the 20-layered and pre-trained CNN models were introduced. Particularly in the 20-layered model, the three fully connected layers were evaluated before the classification layer. Further, the last block layers were applied for the pre-trained model to obtain more specific features from the skin lesion images. Lastly, the weighted ensemble approach was conducted to improve classification performance. The observed results on the HAM10000 dataset indicate that the proposed strategy has improved an accuracy by 0.43% and 2.99% for 20-layered and pre-trained CNNs, respectively. Furthermore, the proposed WELDONNet model outperformed other CNNs with an accuracy of 99.36%. In sum, the proposed model was applicable to be implemented as a reference for skin cancer early detection systems as well as for prospective further research.