An optimized convolutional dense network for skin cancer detection in dermatological diagnostics

Early skin cancer detection is essential because it improves survival rates, enables timely and less invasive treatment, reduces healthcare costs, prevents disease progression, and increases public awareness, ultimately promoting better long-term health and more effective preventive care strategies. However, prevailing approaches face challenges, namely high false-positive rates, overfitting, and reduced accuracy. This work presents a novel approach named convolutional dense network-caracal narwhal optimization algorithm (CDen-Net), for skin cancer detection leveraging input skin images. The process starts by retrieving the skin image selected from the available dataset, followed by pre-processing with a sharpening filter. Next, the skin lesion is segmented employing an adaptive dual-stream hybrid encoder and focal attention decoder network, where the loss functions are adjusted with the proposed focal-cosh loss. Image augmentation is then performed on the segmented images using region cropping, color jittering, and dihedral transformation. Various features are extracted from the augmented images, and diagnosis of skin cancer is carried out utilizing the convolutional dense network (Conv-DenNet). Finally, the network's parameters are fine-tuned for optimal performance using the proposed caracal narwhal optimization algorithm. Furthermore, the developed CDen-Net achieved a false positive rate of 4.532%, an accuracy of 96.382%, and a true positive rate of 97.592%.