Application of Hybrid U-Net Model in Skin Lesion Boundary Segmentation

Tóm tắt

The diagnosis of dermatological diseases through dermoscopy images has been a subject of extensive research aimed at assisting medical experts in the early detection of life-threatening lesions, such as skin cancer, especially within the constraints of limited healthcare resources. Recent advancements in deep learning models have been highly regarded for their capacity to automatically extract complex features and process large-scale datasets with high precision. However, current direct diagnostic approaches applied to original dermoscopy images often suffer from significant error margins. These inaccuracies are primarily attributed to various noise factors, including the presence of hair, medical rulers, air bubbles, and uneven lighting conditions that obscure the lesion boundaries. To enhance the robustness and accuracy of skin lesion diagnosis, this project proposes a collaborative model leveraging the Hybrid U-Net architecture. This architecture integrates a ResNet-34 backbone as an encoder to capture multi-scale spatial features and incorporates the concurrent Spatial and Channel Squeeze and Excitation (scSE) attention mechanism. The inclusion of scSE blocks allows the model to recalibrate feature maps by emphasizing informative spatial regions and channel characteristics. This approach ensures more accurate representation of lesion boundaries and effectively eliminates background noise before the final classification stage. Experimental results on the preprocessed HAM10000 dataset demonstrate that the proposed model achieves a mean Dice score of 0.9466 and IoU of 0.9051, outperforming all compared architectures in both mean score and standard deviation — indicating a favorable accuracy-stability trade-off for clinical dermoscopy segmentation.