关键词:
深度学习
语义分割
卷积神经网络
U-Net
计算机视觉
摘要:
乳腺癌是女性常见恶性肿瘤之一,严重威胁女性健康。在乳腺癌手术中,准确测量乳房和切除组织的体积对于手术规划和乳腺重建至关重要。本研究构建了一个基于深度学习的分割框架,用于乳房和切除组织的MRI体积分割。该框架不仅展现了良好的分割精度,还涵盖了偏置场矫正这一关键预处理步骤。我们收集并手动标注了一个包含47例患者的MRI数据集,这些数据涵盖了不同的年龄、乳房大小和四类采集参数。在交叉验证中,U-Net网络在全乳分割和切除组织分割任务中表现最佳,平均Dice系数分别为96.54和92.37。在测试集上,U-Net网络同样展现了优异的分割效果,平均Dice系数分别为94.23和84.53。实验结果表明,所提出的框架能够精准且高效地量化全乳体积和切除组织体积,为临床乳腺手术提供数据支持。Breast cancer is one of the most common malignant tumors in women, which seriously threatens women’s health. In breast cancer surgery, accurate measurement of breast and excised tissue volume is critical for surgical planning and breast reconstruction. This study developed a deep learning based segmentation framework for MRI volume segmentation of breasts and excised tissues. This framework not only demonstrates good segmentation accuracy, but also covers the key preprocessing step of bias field correction. We collected and manually annotated an MRI dataset containing 47 patients, covering different ages, breast sizes, and four types of acquisition parameters. In cross validation, the U-Net network performed the best in both whole milk segmentation and excised tissue segmentation tasks, with average Dice coefficients of 96.54 and 92.37, respectively. On the test set, the U-Net network also demonstrated excellent segmentation performance, with average Dice coefficients of 94.23 and 84.53, respectively. The experimental results indicate that the proposed framework can accurately and efficiently quantify total breast volume and excised tissue volume, providing data support for clinical breast surgery.