Abstract / Summary
Abstract As the global burden of skin cancer continues to rise, early and accurate diagnosis has become increasingly critical for timely intervention and reducing mortality. Recent advances in artificial intelligence, particularly deep learning, have shown considerable potential for automating dermoscopic image analysis and supporting accurate skin lesion diagnosis. However, conventional convolutional neural networks rely on spatially shared kernels that process all image regions uniformly, limiting their ability to adapt to heterogeneous lesion structures and capture subtle diagnostic features. In this study, we propose WaveInvNet, a lightweight multimodal architecture that integrates dermoscopic images with patient metadata through a novel Wavelet-Separated Dynamic Involution (WSDI) block. Using a stationary Haar wavelet transform, the WSDI block separates structural information from fine-grained details and employs these complementary components to generate spatially adaptive involution kernels. The blocks are organized within a four-stage hierarchical encoder to progressively extract frequency-aware visual representations. These features are subsequently concatenated with encoded patient metadata before final prediction. Experimental evaluation on the HAM10000 dataset showed that WaveInvNet outperformed the compared deep learning architectures, achieving an accuracy of 93.63%, an AUC of 98.46%, and an F1-score of 94.81%, while requiring a relatively small number of trainable parameters.