ADDRESSING IMBALANCE ISIC-2019 DATASET IN DERMOSCOPIC PIGMENTED SKIN LESION CLASSIFICATION

Skin cancer, particularly melanoma, is a life-threatening condition that requires early detection and prompt treatment to reduce mortality rates. Dermoscopic images offer a non-invasive and cost-effective method for examining pigmented skin lesions. However, accurate analysis is challenging due to the lack of standardized colours, image capture settings, and artefacts. Deep learning models, such as Convolutional Neural Networks (CNNs), have shown promising Computer-Aided Diagnosis (CAD) results by automatically extracting features from medical images. Nonetheless, the performance of these models heavily relies on the quality and balance of the training dataset. This study addresses the imbalance issue within the ISIC-2019 dataset, which contains dermoscopic images of pigmented skin lesions. Three pre-trained CNNs (Inception-v3, Xception, and DenseNet-201) were chosen to implement the augmentation scenario. The dataset underwent several preprocessing steps, including duplicate detection, data cleaning, and resizing. Additionally, data augmentation techniques were applied to balancing the distribution of images across different classes. Experimental results demonstrated the effectiveness of the proposed method in improving the classification performance of the pre-trained CNNs. These findings underscore the significance of dataset preparation and augmentation techniques in overcoming challenges posed by imbalanced datasets. The results validate the effectiveness of data preprocessing and augmentation techniques in achieving higher classification accuracy. © 2024 ICIC International. All rights reserved.