DenseNetWithMFA: Enhancing Skin Disease Diagnosis with Multi-Feature Fusion

The diagnosis of skin diseases poses a significant medical challenge in developing countries, with millions of people dying each year from complications related to skin diseases. Previous skin disease detection models have had issues such as low accuracy and overly theoretical constraint indicators [1]. In early models, skin lesions were not clearly identified, making advanced, efficient, and effective skin disease detection essential. To overcome the limitations of existing models, this paper proposes a skin disease diagnosis method called DenseNetWithMFA (Multi-feature attention). In the initial phase, skin disease images were collected for training from Kaggle datasets and research papers. Region growing segmentation augmented with edge refinement was used for preprocessing, ensuring accurate lesion localization before DenseNetWithMFA training. The test image datasets include the Skin Conditions Image Dataset (SCID) and Web-scraped Skin Image Dataset(WSID), and were validated on the test set to diagnose the input skin condition. DenseNetWithMFA achieved recognition accuracies of 99.8% and 97.1% on high-quality WSID and SCID datasets, respectively, surpassing the performance of existing methods.