DermMultiNet: Deep-Learning Based Classification of Skin Lesions from Dermascopic Images

Skin cancer presents a significant global health challenge, making timely detection crucial for improving patient outcomes. Traditional diagnostic methods often demand extensive expertise and time, emphasizing the need for automated, highly accurate solutions. This study introduces DermMultiNet, a novel hybrid Deep Learning (DL) for skin lesion classification. The model is designed to capture local and global features simultaneously while incorporating multi-frequency information extraction. DermMultiNet integrates a Dilated Residual Block (DRB) and a Deformable Separable Depth Block (DSDB) for comprehensive feature extraction, capturing high-level and mid-level image details. The Efficient Multi-Feature Transformer Block (EMFTB) enhances the model's ability to learn multi-frequency information and accurately delineate disease boundaries. The final classification is achieved through a feed-forward network (FFN) classifier. This approach achieves an accuracy of 84.03%, precision of 83.27%, recall of 84.03%, and an F1-score of 83.57% on the HAM10000 dataset.