关键词: 医图分割, 深度学习, 注意力U型网络, 对比学习, 特征融合

Abstract: CT Image and Magnetic Resonance Image (MRI) were widely used in medical image depth segmentation. However, traditional segmentation methods were affected by tumor boundary blurring and structural complexity, and the impact of adversarial samples on the differentiation ability of the segmentation model was ignored, leading to challenges in achieving optimal segmentation results. To address these issues, a 3D medical image segmentation model with adversarial sample-embedded attention U-network learning was proposed. Adversarial samples were constructed through sample transformation, and tumor feature information from medical images was extracted. Low-dimensional feature screening and high-dimensional feature fusion modules were introduced to purify the tumor-distinguishable feature representation. The entire network was trained using a combined loss function based on cross-entropy, Dice loss, and comparison loss to obtain a segmentation model rich in discriminative features. The Dice similarity coefficients for the method on Nerve Sheath Tumor (NST) and Automated Cardiac Diagnosis Challenge (ACDC) reached 88.14% and 91.75%, respectively, which represented improvements of 1.26 percentage points and 2.48 percentage points compared to the nnU-Net method. Experimental results showed that the proposed method effectively improved the performance of 3D medical image segmentation when the tumor boundary was blurred.

Key words: Medical image segmentation, deep learning, attention U-Net, contrastive learning, feature fusion