A deep learning-based auto-segmentation system for organs-at-risk on whole-body computed tomography

https://www.sciencedirect.com/science/article/pii/S0167814021062174

看了这篇论文，结果做的很细致、很详实，不过还是感觉有点虚~

首先讲的是设计理念

1、先通过2.5D Unet网络来分割thorax，

2、根据预测的thorax中的Lungs和一些经验边界值来划分body part

3、根据划分的部位来分别预测。

使用的方法

head neck	Uanet（deepvoxel自己的）
thorax	2.5D Unet
Abdomen+Pelvis	3D Unet

但是里面完整的流程很值得借鉴：

1、数据集的建议

2、勾画指导

head neck	CT-based delineation of organs at risk in the head and neck region: DAHANCA, EORTC, GORTEC, HKNPCSG, NCIC CTG, NCRI, NRG Oncology and TROG consensus guidelines
thorax	Consideration of dose limits for organs at risk of thoracic radiotherapy: atlas for lung, proximal bronchial tree, esophagus, spinal cord, ribs, and brachial plexus
Abdomen	Upper abdominal normal organ contouring guidelines and atlas: a Radiation Therapy Oncology Group consensus
Pelvis	Pelvic normal tissue contouring guidelines for radiation therapy: a Radiation Therapy Oncology Group consensus panel atlas

 3、器官部位划分

head neck	28 OARs	1、brachial plexus, 2、brain stem, 3、constrictor naris, 4、ears (left and right), 6、eyes (left and right), 8、hypophysis, 9、larynx, 10、lenses (left and right), 12、mandible, 13、optical chiasm, 14、optical nerves (left and right), 16、oral cavity, 17、parotids (left and right), 19、submandibular glands (left and right), 21、spinal cord, 22、sublingual gland, 23、temporal lobes (left and right), 25、thyroid, 26、temporomandibular joints (left and right), 28、trachea.
thorax	6 OARS	1、esophagus, 2、heart, 3、lungs (left and right), 5、spinal cord, 6、trachea
Abdomen+Pelvis	16 OARS	1、small bowel, 2、large bowel, 3、duodenum, 4、gallbladder, 5、kidneys (left and right), 7、liver, 8、pancreas, 9、spinal cord, 10、spleen, 11、stomach, 12、bladder, 13、femur heads (left and right), 15、rectum, 16、bowel bag.

 实际上50个是有交集的

 4、还有一些比较的参数，可以作为baseline参考

 

里面有一些对dose的分析，不懂就跳过了，论文写得很详尽，但对其他网络的比较就不一定公正了。

uanet直接我也接触过，确实是比较优秀的网络，主体思想很棒。because RCNN based segmentation models tend to require more data so that the proposal network can generate reliable regions of interest;好像是有这个说法

Varian Eclipse and MIM用的是基于形变配准的atlas分割么，不知道啊~

为什么不用全身数据：

1、是获取不到

2、没办法一起训练，论文说超过40个显卡就吃不消了

推理用的是1080ti

论文说brachial plexus, hypophysis, optic chiasm, and sublingual gland效果比较差是因为器官小和对比度不够，我个人觉得类别不均也是个也主要的问题，不过也能归结于器官小吧。

总的来说是一片比较实在的论文吧
 

另外测试了一下https://irvine.deep-voxel.com/#/dicoms

感官上还可以，不过感觉没有把最好的结果展示出来，准确度和边界光滑程度都不是特别好。

 最后点题了一下transfer learning or few-shot有机会确实要学习一下