بهبود بخش‌بندي تصاوير پزشكي با تقويت گيت‌هاي توجه در معماري U-Net

Medical image segmentation represents one of the fundamental challenges in the fields of image processing an‎d deep learning, playing a pivotal role in disease diagnosis, treatment planning, an‎d patient monitoring. Convolutional neural networks based on the U-Net architecture, owing to their symmetric encoder-decoder structure an‎d skip connections, have become the de facto gold stan‎dard in this domain. Nevertheless, these networks face inherent limitations, including the lack of focused attention on salient image regions, sensitivity to complex backgrounds, an‎d insufficient utilization of structural information such as edges an‎d textures, all of which can compromise segmentation accuracy, particularly at lesion boundaries. This research aims to enhance the performance of medical image segmentation by designing, implementing, an‎d comparing four distinct U-Net-based architectures. In addition to the stan‎dard U-Net (serving as the baseline) an‎d Attention U-Net (incorporating stan‎dard attention mechanisms), two novel architectures are proposed: FEA-U-Net (Feature Enhanced Attention U-Net) an‎d EAG-U-Net (Edge-Aware Gated U-Net). The former leverages trainable convolutional layers to extract textural features, while the latter employs classical edge detection filters, specifically Sobel an‎d Laplacian operators, to capture structural information, both integrated with attention mechanisms. The dataset utilized in this study is a composite of two CT scan image collections from patients diagnosed with COVID-19, comprising a total of 929 images. Following preprocessing an‎d integration, the data were eva‎luated within a binary segmentation framework, distinguishing infected regions from healthy tissue. Experimental results on two independent test sets demonstrate that both proposed architectures substantially outperform the baseline networks. In the first eva‎luation on 110 samples, FEA-U-Net achieved a Dice coefficient of 0.7499, representing an improvement of 24.4% over stan‎dard U-Net an‎d 12.1% over Attention U-Net. Concurrently, this network reduced the loss function by 35.2% an‎d increased sensitivity by 11.8%, establishing a superior balance between precision an‎d comprehensive lesion detection. In the second eva‎luation on 200 samples, EAG-U-Net, employing classical edge extraction, attained a Dice coefficient of 0.8293, outperforming the baseline architectures. The sensitivity of this network reached 0.6812, reflecting improvements of 4% an‎d 11% over the stan‎dard U-Net an‎d Attention U-Net, respectively. A particularly noteworthy finding is the remarkable stability of the proposed networks, evidenced by the reduction in the stan‎dard deviation of the Dice coefficient to 0.0102 for FEA-U-Net, underscoring the reproducibility an‎d clinical reliability of these methods. Qualitative analysis further corroborates these findings, revealing that the proposed networks generate more precise segmentation maps, especially in cases with indistinct boundaries an‎d small lesions, thereby offering greater diagnostic value in clinical settings. Moreover, the comparison between the two proposed approaches indicates that the classical edge-based filtering strategy (EAG-U-Net), despite its simplicity an‎d lack of trainable parameters, delivers competitive performance relative to the trainable counterpart (FEA-U-Net), making it a computationally efficient option for resource-constrained applications. In conclusion, the findings of this study affirm that integrating structural information—specifically edges an‎d textures—with attention mechanisms constitutes an effective strategy for overcoming the limitations of baseline segmentation networks. The proposed architectures, characterized by high accuracy, robust performance stability, an‎d satisfactory generalizability, hold significant promise as efficient tools for computer-aided diagnosis an‎d clinical decision support systems.

بهزاد نظري

جعفر قيصري , فروغ السادات طباطباء