Soft Attention Improves Skin Cancer Classification Performance

In clinical applications, neural networks must focus on and highlight the most important parts of an input image. Soft-Attention mechanism enables a network to achieve this goal. This paper investigates effectiveness in deep architectures. The central aim is boost value features suppress noise-inducing features. We compare performance VGG, ResNet, Inception ResNet v2 DenseNet architectures with without mechanism, while classifying skin lesions. original when coupled outperforms baseline [16] by 4.7% achieving precision 93.7% HAM10000 dataset [25]. Additionally, coupling improves sensitivity score 3.8% compared [31] achieves 91.6% ISIC-2017 [2]. code publicly available at github (https://github.com/skrantidatta/Attention-based-Skin-Cancer-Classification).