基于深度学习的脑肿瘤图像分割方法研究
Research on Brain Tumor Image Segmentation Methods Based on Deep Learning

脑肿瘤是严重威胁人类健康的疾病之一，因此脑肿瘤精确分割在临床诊疗中非常重要。然而在实际医疗诊断中要面临瘤体体积、位置难以确定，形状各异的问题，致使脑肿瘤分割任务面临着分割结果精确度不够高、训练计算量较大以及分割时间久的问题。近年来，U-Net凭借其简洁的架构和优秀的性能成为解决医学图像分割任务的主流模型，但其也存在特征提取能力有限、局部感受野有限、对复杂目标的适应性不够和上下文信息利用不足等问题。本文针对以上问题，提出一种U-Net网络改进方案。通过将ResNet50替换U-Net网络的主干特征提取网络，并在跳跃连接和解码器上采样中引入SE-Net模块，提升网络的表现力和鲁棒性，选用Dice Loss与Binary Cross Entropy (BCE) Loss结合的损失函数，主要作用是优化深度学习模型在脑肿瘤分割任务中的性能。在公开的BraTS2021数据集上进行了大量实验，结果显示本文模型的Dice评价指标达到0.856，相较于U-Net指标改善了4.3%。同时与其他代表性方法对比也表现出了良好的脑肿瘤分割性能。
Brain tumors are one of the diseases that pose a serious threat to human health, making precise brain tumor segmentation critically important in clinical diagnosis and treatment. However, in practical medical diagnostics, challenges such as difficulties in determining tumor volume, location, and the variability in tumor shapes lead to issues in brain tumor segmentation, including insufficient accuracy, high computational costs, and prolonged segmentation times. In recent years, U-Net has become a mainstream model for medical image segmentation due to its simple architecture and excellent performance. However, it also suffers from limitations such as limited feature extraction capabilities, restricted local receptive fields, inadequate adaptability to complex targets, and insufficient utilization of contextual information. To address these issues, this paper proposes an improved U-Net network. By replacing the backbone feature extraction network of U-Net with ResNet50 and introducing the SE-Net module in skip connections and decoder upsampling, the network’s expressiveness and robustness are enhanced. Additionally, a combined loss function of Dice Loss and Binary Cross Entropy (BCE) Loss is employed to optimize the performance of the deep learning model in brain tumor segmentation tasks. Extensive experiments conducted on the publicly available BraTS2021 dataset show that the proposed model achieves a Dice evaluation metric of 0.856, representing a 4.3% improvement over the original U-Net. Furthermore, the model demonstrates superior brain tumor segmentation performance compared to other representative methods.