基于多尺度特征感知的胸腔图像危及器官分割

医学图像危及器官自动分割是计算机辅助诊断中的重要组成部分,对辅助医生高质高效完成放射治疗有着极其重要的作用。胸腔CT图像对比度低,且各器官之间重叠交错、边界模糊,使得危及器官的精确分割具有较大的挑战性。提出一种多尺度特征感知的编码-解码网络模型(FA-Unet),实现胸腔CT图像危及器官的分割。针对胸腔中四类器官大小差异的问题,首先构建了输入感知模型,提取图像中各器官的多尺度特征。为了弥补编码与解码之间的语义鸿沟,在解码-编码中融入改进的inception模块。用空间金字塔卷积(ESP)与金字塔池化(PSP)模块代替传统的串行卷积运算,使得网络模型更为轻量化,在一定程度上缓解数据量不足带来的过拟合问题。采用一种联合Dice系数与交叉熵的损失函数训练分割网络,可解决胸腔CT图像中类别不平衡的问题。最后,在2019年ISBI发布的SegTHOR数据集上验证模型的有效性,该数据集共包括40例肺癌或霍奇金淋巴瘤患者的胸腔CT图像7 390张。实验结果表明,胸腔CT图像各器官分割的Dice系数分别为食道0.793 2、心脏0.935 9、气管0.854 9、主动脉0.889 0,Hausdorff距离分别为食道1.420 7、心脏0.212 4、气管0.627 3、主动脉0.887 0。结果表明,与同类型分割网络相比,模型可获得较好的分割性能,尤其在小目标器官的分割上取得竞争性优势。

Segmentation of Thoracic Image Organs at Risk Based on Multi-Scale Feature-Aware

Automatic segmentation of organs at risk (OARs) in medical images is an essentialconstituent of computer-aided diagnosis, and it plays a vital role in assisting doctors to completeradiotherapy with high quality and efficiency. There are some challenges in the accuratesegmentation of OARs for thoracic CT images, including low intensity contrast, different organswith interlaced and overlap regions, and different structure without clear boundaries. In this paper,a multi-scale feature-aware encoding-decoding network (FA-Unet) was proposed to segmentOARs in thoracic CT images.To address the problem of the size difference among fourkinds of organs in the thoraciccavity, an input-aware module was designed to extract multi-scale features in four types of organs. In order to bridge the semantic gap between theencoding and decoding layers, the modified inception module was introduced to long-range skipconnections between the encoding part and the decoding part in our architecture. Furthermore,we replaced the traditional serial convolution operation with the efficient spatial pyramid(ESP) andpyramid spatial pooling (PSP) modules to make our network more lightweightand avoide over-fitting caused by insufficient data effectively. We formulated a novel lossfunction by combining Dice coefficient and cross entropy to train our network to resolve theclass imbalance in thoracic CT images. Finally, we evaluated the effectiveness of our model onthe SegTHOR data set released by ISBI in 2019, and the dataset includes 7 390 thoracic CTimages of 40 patients with lung cancer or Hodgkin's lymphoma. Experimental results showedthat Dice coefficient of each organ in thoracic CT image was 0.793 2 of esophagus, 0.935 9of heart, 0.854 9 of trachea and 0.889 0 of aorta. Hausdorff distances was 1.420 7 of esophagus,0.212 4 of heart, 0.627 3 of trachea and 0.887 0 of aorta. Experimental results verified that ourproposed model outperformed other state-of-the-arts on the segmentation results of OARs andachieved very competitive performance on small target organs.