针对目前磁共振脑影像上脑白质高信号区域的自动分割存在分割精度较低和细小病灶易漏识等问题,提出一种融合注意力和Inception的U-Net分割模型。在U-Net的编码阶段加入Inception模块以增加网络宽度,使其具有多尺度特征的提取能力,并加入注意力模块以增强网络对分割目标的关注度,两者的加入和融合可以有效提升网络的特征提取和表达能力。同时,在解码阶段的每一个卷积层上增加残差连接,可以提高网络的优化速度。此外,针对样本不均衡易导致分割结果中假阴性过多的问题,采用具有均衡调节能力的Tversky损失函数优化网络训练。实验结果表明,所提方法能够较好地分割脑白质的高信号区域,特别是小病灶区,且各项分割指标优于多个对比方法。

Because the brain edema has a crucial impact on morbidity and mortality, it is important to develop a noninvasive method to monitor the process of the brain edema effectively. When the brain edema occurs, the optical properties of the brain will change. The goal of this study is to access the feasibility and reliability of using noninvasive near-infrared spectroscopy (NIRS) monitoring method to measure the brain edema. Specifically, three models, including the water content changes in the cerebrospinal fluid (CSF), gray matter and white matter, were explored. Moreover, these models were numerically simulated by the Monte Carlo studies. Then, the phantom experiments were performed to investigate the light intensity which was measured at different detecting radius on the tissue surface. The results indicated that the light intensity correlated well with the conditions of the brain edema and the detecting radius. Briefly, at the detecting radius of 3.0 cm and 4.0 cm, the light intensity has a high response to the change of tissue parameters and optical properties. Thus, it is possible to monitor the brain edema noninvasively by NIRS method and the light intensity is a reliable and simple parameter to assess the brain edema.