Revolutionizing skin lesion classification using novel deep learning models

Abstract

Skin cancer is a significant global health concern, and timely, accurate diagnosis is crucial for improving patient outcomes. Recent advancements in deep learning have revolutionized automated skin cancer classification, offering powerful tools for dermatology. This thesis presents three novel approaches to enhance skin lesion classification accuracy using deep learning techniques. The first method utilizes a deep feature fusion network (DFF-Net), enhanced by squeeze-and-excitation attention, to tackle the classification task. Images are initially cropped using a contour-based approach, and features are then extracted from these cropped images using Xception and BiT-M models. These features are fused and processed through a squeeze-and-excitation attention module, followed by pooling and fully connected layers for final classification. A combined loss function is introduced to further boost performance. The second method introduces a novel edge-guided dual-branched network (EDB-Net) for classification. One branch processes edge maps of dermoscopic images, while the other focuses on the original images. Both branches use a DenseNet-based backbone with dense and transition blocks to extract features. The second branch incorporates an additional attention module to highlight salient regions in the images, capturing more relevant features. A novel strategy modulates attention maps using previously extracted edge features, creating edge-guided features that refine the representations. The final stage aggregates both attended and edge-guided features, which are then flattened and fed into dense layers for classification. The third method employs a Dempster-Shafer theory based classifier combination network (DSCC-Net) that combines the outputs of three attention-based convolutional neural networks. The base learners — ResNet50, InceptionRes- NetV2, and DenseNet201 — are each enhanced with soft attention modules to improve focus on relevant features. These base learners generate confidence scores which are then combined using the popular Dempster-Shafer theory. The three proposed models are evaluated on three challenging skin cancer datasets: HAM10000, ISIC 2016, and ISIC 2017. These approaches exhibit superior performance compared to many recent methods in the literature, achieving state-of-the-art accuracy in skin lesion classification and demonstrating significant advancements in this domain.