An Attention-based Hybrid Deep Learning Approach for Skin Lesion Classification

Skin cancer is the most frequent type of cancer worldwide in recent years. It affects more than a million people annually, a large number of whom are elderly. Because of incredibly fine-grained differences of skin lesion, it is difficult to classify distinct forms of skin lesions using images. An automated system ought to be able to classify these skin lesions with reasonable accuracy. Preprocessing steps like Blackhat filter, Median filter including resizing, augmentation and labeling were performed before uploading the images to the system. For the model, a sizable number of images must be used for training to yield the required results. This model is trained on 2 different dataset: HAM10000 (Human Against Machine) with 10,015 images on skin lesion and ISIC(The International Skin Imaging Collaboration). Here A hybrid CNN model combined with soft attention was developed for effective multi-class classification. Here, ConvNet (A customized CNN) and DenseNet-201 is used to the proposed hybrid model. In previous several models were implemented and it is found that, by combining the best characteristics of both architectures, the DenseNet-ConvNet model has the potential to improve feature learning by fusioning low level and high level features, increase accuracy, and even increase parameter efficiency. The model yields an accuracy of 96.13% for HAM-10,000 and 92.36% for ISIC dataset, attaining f1-score of 96% on HAM-10,000 and 92.37% on ISIC.