基于深度学习构建微创肝切除术关键解剖结构识别模型的应用价值

Application value of major anatomical structure recognition model of minimally invasive liver resection based on deep learning

目的 探讨基于深度学习构建微创肝切除术关键解剖结构识别模型的应用价值.方法 采用回顾性描述性研究方法.收集2019年1月至2023年4月在南方医科大学珠江医院保存完整的31个腹腔镜肝左外叶切除术视频资料.由2位肝脏外科医师筛选出包含肝左外叶肝蒂、肝左静脉的视频片段,对视频进行质控、筛选、抽帧后,标注图像中的关键解剖结构.所有图像经过预处理后,输送至DeepLab v3+神经网络框架进行模型训练.观察指标:(1)视频标注及分类情况.(2)人工智能解剖识别模型测试结果.正态分布的计量资料以(-x)±s表示,计数资料以绝对数表示.结果 (1)视频标注及分类情况.31个视频共标注4 130帧图像,其中仅有肝左外叶肝蒂的标注图像2083帧,仅有肝左静脉的标注图像1 578帧,同时有肝左外叶肝蒂和肝左静脉的标注图像469帧.(2)人工智能解剖识别模型测试结果.4种场景(干净场景、血渍场景、器械部分遮挡场景、暴露面积小的场景)下,模型成功识别肝左外叶肝蒂或肝左静脉.模型对以上视频图像的解剖识别速度均>13帧/s.模型在对仅有肝左外叶肝蒂的图像进行解剖识别时,Dice系数、交并比、准确度、灵敏度、特异度分别为0.710± 0.110、0.560±0.120、0.980±0.010、0.640±0.030、0.980±0.010;在对仅有肝左静脉的图像进行解剖识别时,上述指标分别为 0.670±0.180、0.530±0.200、0.980±0.010、0.600±0.040、0.990±0.010;在对同时有肝左外叶肝蒂和肝左静脉的图像进行解剖识别时,上述指标分别为0.580±0.180、0.430±0.190、0.980± 0.010、0.580±0.020、0.990±0.010.结论 基于深度学习构建微创肝切除术关键解剖结构识别模型可用于识别肝脏肝蒂和静脉结构.

Objective To investigate the application value of major anatomical structure recognition model of minimally invasive liver resection based on deep learning.Methods The retrospective and descriptive study was conducted.The 31 surgical videos of laparoscopic left lateral sectionectomy performed in Zhujiang Hospital of Southern Medical University from January 2019 to April 2023 were collected.Video clips containing the surgical procedure of left lateral lobe liver pedicle and left hepatic vein were screened by 2 liver surgeons.After quality control,screening and frame extraction,the major anatomical structures on the images of these clips were annotated.After pre-processing,these images were transported to the DeepLab v3+neural network framework for model training.Observation indicators:(1)video annotation and classification;(2)results of arti-ficial intelligence anatomical recognition model testing.Measurement data with normal distribution were represented as Mean±SD,and count data were described as absolute numbers.Results(1)Video annotation and classification.A total of 4130 frames of images were annotated in the 31 surgical videos,including 2083 frames of annotated images for the left lateral lobe liver pedicle,1578 frames of annotated images for the left hepatic vein and 469 frames of annotated images for both the left lateral lobe liver pedicle and left hepatic vein.(2)Results of artificial intelligence anatomical recognition model testing.In four application scenarios(clean scene,bloodstain scene,partially obstruction by instrument scene,and small exposed area scene),the model was able to successfully recognize the left lateral lobe liver pedicle and left hepatic vein,with a recognition speed for anatomical markers>13 frames/s.When performing anatomical recognition on images with only the left lateral lobe liver pedicle,the Dice coefficient,intersection over union,accuracy,sensitivity and specificity of the model were 0.710±0.110,0.560±0.120,0.980±0.010,0.640±0.030,and 0.980±0.010,respectively.The above indicators of the model were 0.670±0.180,0.530±0.200,0.980±0.010,0.600±0.040,and 0.990±0.010 when performing anatomical recognition on images with only the left hepatic vein,and 0.580±0.180,0.430±0.190,0.980±0.010,0.580±0.020,and 0.990±0.010 when per-forming anatomical recognition on images with both the left lateral lobe liver pedicle and left hepatic vein.Conclusion The major anatomical structure recognition model of minimally invasive liver resection based on deep learning can be applied in identifying liver pedicle and hepatic vein.