Dense-par-AttNet: An Attention Based Deep Learning Model For Skin Lesion Classification By Transfer Learning Approach

The classification of dermatoscopy images is of great significance, especially in the case of skin cancer, as the chance of survival degenerates with the passage of time. Yet, detection of a particular class of skin cancer has become a challenge in medical diagnosis due to the close resemblance among various lesions. As existing Computer-Aided Diagnosis (CAD) methods that optimize deep networks fail to perform up to the mark due to fuzzy boundaries, low contrast and limited training sets, this paper proposes a new attention-based transfer learning approach for the classification of skin lesions. In this method, pre-trained DenseNet-201 has been imported in addition to a spatial attention-based CNN network. The extracted feature of both networks are merged together to make the optimum prediction. The experimental results demonstrate the considerable performance of 82.576% overall accuracy for the HAM10000 dataset. The proposed system has a great prospective to be applied in hospitals to help dermatologists make accurate decisions in the case of skin lesion classification.