Ensembling Transfer Learning Frameworks for Effective Lightweight Skin Disease Detection

Skin disease diagnosis is challenging due to various skin lesion texture and detection challenges. Processing complex skin lesions with detailed features, like limited training datasets, poor background contrast, the presence of artifacts, and ambigu- ous boundaries, remains a challenge for traditional deep learning- based techniques. Moreover, these approaches require multiple parameters in deep architecture, resulting in higher computing resource consumption, poor generalization, and overfitting issues. This paper leverages the ensembling of two popular transfer learning frameworks for skin disease detection. It employs an amalgamation of popular deep learning architectures, such as Visual Geometry Group (VGG)-16 and MobileNet, to classify skin diseases. The proposed model is lightweight and can de- tect skin diseases efficiently. The performance of the proposed framework is assessed in light of the well-known HAM10000 and MNIST datasets. TThe suggested model's relative performance is evaluated against several conventional methods. We show that the ensemble framework exhibits superior performance in comparison to other deep neural network models.