Attention-Based Convolutional Neural Network Model for Skin Cancer Classification

Early detection of skin cancer (SC) is paramount for effective treatment. Although convolutional neural networks (CNN) have facilitated automated learning of high-level features from dermoscopic images, many prior studies relied on separate feature extraction and selection stages combined with machine learning classifiers, resulting in increased model complexity and a higher risk of overfitting. Furthermore, sensitivity of conventional CNNs to subtle lesion changes can be challenging without inherent mechanisms to emphasize salient image regions, ultimately hindering classification performance. Moreover, the ambiguous nature of these models often leaves dermatologists uncertain about the reasoning behind predictions, which limits their adoption in clinical settings. Therefore, we propose an efficient dual attention-based ResNet50 (EDA-ResNet50) model that effectively captures vital details in lesion areas. EDA-ResNet50 integrates a multi-scale feature representation block to effectively capture features across various spatial scales. Additionally, a unique efficient dual attention module is introduced, amalgamating a lightweight convolution-based recalibration unit with channel and spatial attention blocks to enhance information flow. Gradient-weighted class activation mapping is exploited to understand the EDA-ResNet50 model decisions. The proposed system classifies dermoscopic images into benign and malignant categories with accuracy of 93.18%, offering confidence for dermatologist. EDA-ResNet50 outperformed several CNN techniques. Unlike existing studies, EDA-ResNet50 streamlined architecture presents a compelling solution for automated SC classification, combining high diagnostic accuracy with low computational overhead. Moreover, EDA-ResNet50 holds significant potential for application to other image classification tasks. EDA-ResNet50 presents a new frontier in the intersection of deep learning and healthcare, empowering medical professionals with efficient diagnostic tools.