Q349 : Skin cancer recognition in dermatoscopic images using attention mechanisms
Thesis > Central Library of Shahrood University > Computer Engineering > MSc > 2026
Authors:
Abstarct: Abstract
The development of computer-aided diagnosis (CAD) systems for clinical skin imaging faces persistent structural challenges, such as the presence of physical artifacts (hair lines, gel bubbles, and severe lighting variations) and a severe statistical class imbalance. Traditional computer vision approaches suffer from the rigidity of handcrafted feature engineering, while reference deep convolutional neural networks (CNNs), relying strictly on uniform local receptive fields, lack the capability to isolate background noise from genuine pathological signals. To overcome these theoretical limitations, this thesis proposes an efficient multi-scale hybrid architecture that achieves pure and stable skin lesion classification, specifically differentiating malignant melanoma from benign samples without requiring destructive physical pre-processing algorithms. The proposed architecture is built upon a hierarchical convolutional backbone equipped with dilated filters. To simultaneously preserve fine-grained and global clinical details, two independent computational branches are extracted from the intermediate (Block 2) and deep (Block 4) levels of the network. The extracted tensors are dynamically weighted through Squeeze-and-Excitation (SE) channel attention blocks to suppress noise frequencies from hair and specular reflection while magnifying the lesion's structural attributes. Next, the concatenated tensor enters a Dense Spatial Refinement module. Utilizing three parallel branches with diverse dilation rates spatial relationships and the irregular biological geometry of the lesion across multiple scales. The refined maps are then consolidated via a Fuse block and channeled into the final classification head. Experimental evaluation of this model on the reference international ISIC 2020 dataset demonstrates the decisive superiority of the proposed methodology across four key performance metrics. The designed architecture successfully achieves an overall accuracy of 96%, representing a significant statistical leap over recent state-of-the-art methods. Furthermore, by incorporating a weighted cross-entropy loss function, the model mitigates the acute class imbalance challenge, securing a recall of 91.74% and a precision of 90.13% specifically for the aggressive melanoma class. The stability of the F1-Score above 90% for both classes validates the robustness of the model’s decision boundaries in controlling Type I and Type II errors, establishing its theoretical and practical efficacy as a highly reliable assistive tool for real-world clinical environments.
Keywords:
#_Dermoscopic Imaging #Melanoma #Multi-scale Convolutional Networks #Channel Attention Mechanism #Dilated Convolution_ Keeping place: Central Library of Shahrood University
Visitor:
Visitor: