International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Enhanced Melanoma Detection: An Improved Deformable DETR Model with Efficient Channel Attention

Authors :

Amirreza Rouhbakhshmeghrazi¹ Shayan Nalbandian² Sheida Shadman³ Mohammad Reza Hassannezhad⁴ Shuyuan Yang⁵ Bo Li⁶

1- School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China 2- School of Software Engineering, Northwestern Polytechnical University, Xi'an, China 3- School of Software Engineering, Northwestern Polytechnical University, Xi'an, China 4- School of Aeronautics Northwestern Polytechnical University Xi’an, China 5- School of Artificial Intelligence Xidian University Xi’an, China 6- School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China

Keywords :

Medical imaging،Vision Transformer،Skin lesion،CNN،Attention Mechanism

Abstract :

Deep learning has significantly advanced healthcare, particularly in medical imaging for disease diagnosis. However, detecting small objects and efficiently extracting features remain challenging. Melanoma detection, which requires high accuracy for effective treatment, is especially affected by these limitations. Current models often struggle with feature extraction and localization, particularly for small or rare lesions.In this paper, we propose an improved Deformable DETR model for skin lesion detection in melanoma. We integrate the Efficient Channel Attention (ECA) module into the ResNet backbone of Deformable DETR. ECA enhances feature extraction by adjusting channel weights, allowing the network to focus on important features and suppress less relevant ones.Our results show significant performance improvements. The model achieves an AP of 53.4%, surpassing the baseline Deformable DETR (52.6%) and other models like Faster R-CNN and RetinaNet. Improvements in AP50, AP75, and recall indicate better detection accuracy and reduced false negatives.This work contributes to more accurate and efficient melanoma detection, crucial for reducing mortality rates. By incorporating ECA into transformer-based frameworks, we highlight the potential of attention mechanisms in medical imaging, offering a promising direction for future research.