| In recent times, skin cancer has emerged as a significant global health challenge. Earlier skin cancer disease detection was performed through automated diagnostic models, which are essential for enhancing survival rates and clinical efficiency. However, the existing methods struggle with inter-class similarity and intra-class variability for different lesion types due to similar visual appearance and identical types exhibit significant natural variations. To address these limitations, Skin Stochastic Depth Feature Reuse Network (SkinSDFRNet) uses 48-layer residual architecture to improve feature extraction through a structured sequence process. Depthwise Dilated Convolution (DDC) extracts the feature to expand the receptive field of the kernel without increasing the number of parameters, which enables the network to extract more complex spatial data. Feature Reuse Block (FRB) is employed to maintain spatial resolution and preserve intermediate activations. This preservation is essential for the model to accurately identify irregular boundaries and complex texture patterns within dermoscopic images. Each FRB consists of four linked convolutional layers, where the final output is the concatenation of the processed feature map and the original block input facilitates previously generated features, which are effectively utilized to improve generalization. Finally, the architecture incorporates Multi-Scale Depthwise Dilated Convolution (MDDC) to integrate with Stochastic Depth strategy. The MDDC unit varies dilation rates to capture both minute cellular textures and broader lesion boundaries. The proposed SkinSD-FRNet achieves 98.64% and 98.02% accuracies for Ham10000 and International Skin Imaging Collaboration (ISIC)-2019 dataset respectively, which is superior than existing methods. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.