探讨AI用于食管癌危及器官自动勾画的可行性

目的 探讨人工智能用于食管癌危及器官自动勾画的几何学和剂量学精度。方法 首先,选择100例食管癌患者基于连心人工智能平台建立图像数据库,每个患者包含5个已经手工勾画的危及器官。然后将其他20例患者CT图像传入人工智能平台,系统自动勾画危及器官作为目标图像,与手工勾画的危及器官进行几何学和剂量学比较。最后,分别比较两种勾画方式所需的时间、体积差异、重合度指标、相似性系数和剂量学差异。结果 自动勾画比手工勾画左肺、右肺、心脏、肝脏和脊髓分别节省时间98.83%、94.55%、84.9%、77.96%和94.15%,两者差异有统计学意义（t=2.27、3.28、4.92、-1.39、0.21,P<0.05）。左肺、右肺、心脏、肝脏和脊髓体积差异分别为（5.58±2.53）cm³、（8.57±4.36）cm³、（0.97±0.34）cm³、（1.47±0.65）cm³和（0.73±0.21）cm³,DSC值为0.78~0.96,DSC>0.7,OR值为0.84~0.97,重合度好。自动勾画与手工勾画危及器官剂量学对比中,各项剂量学指标均符合临床要求,除右肺V₅的剂量学指标的差异有统计学意义（t=0.41,P=0.04<0.05）,左肺,右肺,肝脏,心脏和脊髓的其他剂量学指标的差异均无统计学意义（t=-1.23~3.11,P>0.05）。结论 食管癌危及器官的自动勾画几何精度高,剂量学差异小,时间短。AI在临床中的应用,可以大大提升医师的工作效率。

Feasibility of automatic delineation of OAR in radiotherapy of esophageal cancer utilizing AI

Objective To explore the geometric and dosimetric accuracy of automatic delineation of OAR in radiotherapy of esophageal cancer based on Artificial intelligence(AI).Methods Firstly, one hundred cases of esophageal cancer were selected to establish an image database based on the AI of LINKING MED. Each patient included five OARs that manually delineated. Then, the CT images of the other twenty patients were transferred to the AI, and the system automatically sketched the OARs as the target image, which could compare the geometry and dosimetry with the manually sketched OARs. Finally, the time, volume difference, overlap ratio (OR), dice similarity coefficient(DSC)and dosimetric difference of the two methods were compared respectively.Results Compared with manually sketching, the autosegmention saving time by 98.83%, 94.55%, 84.9%, 77.96% and 94.15% in the left lung, right lung, heart, liver and spinal cord respectively, which were significant differences between the two groups(t=2.27, 3.28, 4.92, -1.39, 0.21, P<0.05). The volume differences of left lung, right lung, heart, liver and spinal cord were (5.58±2.53)cm³, (8.57±4.36)cm³, (0.97±0.34)cm³, (1.47±0.65)cm³ and (0.73±0.21)cm³, DSC values were 0.78~0.96, DSC>0.7, and the OR values were 0.84~0.97, which had a good coincidence. In the comparison of dosimetry between autosegmention and manual, all the dosimetric indexes met the clinical requirements. Except for right lung in V₅, which had significant difference in dosimetric indicators (t=0.41, P=0.04<0.05). There was no significant difference in other dosimetric indicators of left lung,right lung, liver, heart and spinal cord (t=-1.23~3.11, P>0.05).Conclusion The automatic delineation of OAR in esophageal cancer has suggests high geometric accuracy, small dosimetric difference and reduced time. The application of AI in clinical practice might greatly improve the work efficiency of doctors.