Multi-Scale Feature Fusion Framework Based on Attention Integrated Customized DenseNet201 Architecture for Multi-Class Skin Lesion Detection

Skin lesions encompass abnormal skin changes, with some evolving into skin cancer if untreated; hence, early detection is vital. Automatic detection utilizing Deep Learning (DL) techniques can enhance accessibility and affordability of diagnosis, overcoming barriers posed by extensive testing and associated costs. However, in most deep learning (DL) research, the issue of imbalanced class distribution is often overlooked, as researchers frequently rely on transfer learning (TL) with pre-trained models without adequate customization or fine-tuning. Consequently, these models suffer in extracting fine-grained details, and there is an absence of analysis regarding the model's decision-making process. Therefore, this paper presents a TL-based explainable Multi-Scale Feature Fusion Framework (MFFF) to address the above-mentioned challenges. This framework utilizes pre-trained DenseNet201 with three attention mechanisms: Channel Attention (CA), Soft Attention (SA), and Squeeze and Excitation Attention (SEA). The top layers of DenseNet201 are customized and fine-tuned with three separate series of convolutional blocks, each comprising diverse attention blocks to extract relevant features at different scales. Finally, the extracted features from these convolutional blocks are concatenated to produce more fine-grained features. An empirical evaluation on a publicly available dataset demonstrates that our proposed MFFF achieves 97.15% accuracy, outperforming other state-of-the-art research. Furthermore, Gradient Class Activation Map (Grad-CAM) is employed to highlight crucial regions and explore the decision-making process of the black box Multi-Scale Feature Fusion Framework (MFFF), assisting dermatologists in effective decision-making.