An Efficient Skin Cancer Detection and Classification Using DenseNet-121 with Attention Mechanism

Diagnosing skin lesions remains a challenging process in medical image analysis, while dermoscopy has emerged as a contemporary approach for early-stage skin cancer diagnosis.Numerous computerized techniques for categorizing skin malignancies have been presented in prior research studies.However, challenges such as poor contrast lesions and removal of redundant or unnecessary features still exist.This research proposes an Attention Mechanism with Densenet-121 classifier to improve skin cancer classification accuracy.Initially, the publicly available datasets named ISCI 2018, 2019, PH2 and HAM10000 are employed to assess the model's efficacy.Then, color representation across input images is normalized using the Gray World Color Constancy method.Then, the Unet is employed for lesion segmentation to help identify individual skin imaging lesions.Furthermore, the hyperparameters of the DenseNet-121 classifier are finetuned using the Improved Cat Swarm Optimization Algorithm (ICSOA) to enable precise classification.The DenseNet-121 performs automatic feature extraction and classification.The proposed attention mechanism modifies the weights of several feature maps in the channel domain to adaptively select the relevant features.Such selected significant features are fed into the classification layer for efficient classification and classification accuracy.Additionally, the channel attention mechanism combines transition layers and the dense block to fully exploit the channel attention module, without introducing any other parameters.Because of the model's compact and efficient architecture, the over fitting issue in the channel attention network is addressed.The comparative analysis results exhibit the proposed ICSOA-DenseNet121 model's superiority over the existing approaches namely: Insinet, Neutrosophic Features with a Deep Neural Network (N-DNN) and Skin-Net, by achieving higher classification accuracies of 99.87%, 99.83%, 99.72%, 99.55% on the ISIC 2017-2020 datasets, and 99.67% on the PH2 dataset.