[High definition MRI rectal lymph node aided diagnostic system based on deep neural network].

Objective: To investigate the clinical significance of high definition (HD) MRI rectal lymph node aided diagnostic system based on deep neural network. Methods: The research selected 301 patients with rectal cancer who underwent pelvic HD MRI and reported pelvic lymph node metastasis from July 2016 to December 2017 in Affiliated Hospital of Qingdao University. According to the chronological order, the first 201 cases were used as learning group. The remaining 100 cases were used as verification group. There were 149 males (74.1%) and 52 females in the study group, with an average age of 58.8 years. There were 76 males (76.0%) and 24 females in the validation group, with an average age of 60.2 years. Firstly, Using deep learning technique, researchers trained the 12 060 HD MRI lymph nodes images data of learning group with convolution neural network to simulate the judgment process of radiologists, and established an artificial intelligence automatic recognition system for metastatic lymph nodes of rectal cancer. Then, 6 030 images of the validation group were clinically validated. Artificial intelligence and radiologists simultaneously diagnosed all cases of HD MRI images and made the diagnosis results of metastatic lymph node. Receiver operating characteristic (ROC) curve and area under curve (AUC) were used to compare the diagnostic level of them. Results: After continuous iteration training of the learning group data, the loss function value of artificial intelligence decreased continuously, and the diagnostic error decreased continuously. Among the 6 030 images of verification group, 912 images were considered to exist metastatic lymph nodes in radiologists' diagnosis and 987 in artificial intelligence diagnosis. There were 772 images having identical diagnostic results of lymph node location and number of metastases with the two methods. Compared with manual diagnosis, the AUC of the intelligent platform was 0.886 2, the diagnostic time of a single case was 10 s, but the average diagnostic time of doctors was 600 s. Conclusion: The HD MRI lymph node automatic recognition system based on deep neural network has high accuracy and high efficiency, and has the clinical significance of auxiliary diagnosis.目的： 探讨基于深度神经网络的高分辨MRI直肠淋巴结识别系统的临床应用价值。 方法： 选取青岛大学附属医院2016年7月至2017年12月术前行盆腔高分辨MRI扫描且报告中明确有盆腔淋巴结转移的直肠癌患者301例，按照就诊时间顺序前201例作为学习组，后100例作为验证组。学习组男性149例，女性52例，平均年龄58.8岁；验证组男性76例，女性24例，平均年龄60.2岁。首先，利用深度学习技术及学习组的12 060张淋巴结高分辨MRI图像，在卷积神经网络下进行训练，模拟影像科医师的判断过程，从而建立了直肠癌淋巴结转移的人工智能自动识别系统。然后，对验证组的6 030张淋巴结高分辨MRI图像进行临床验证，人工智能和影像科医师同时对淋巴结转移情况作出诊断，利用受试者工作特征曲线比较两者的诊断水平。 结果： 经过对学习组数据的不断迭代训练，人工智能的损失函数值不断降低，诊断误差不断降低。验证组的6 030张图像，人工诊断共912张存在淋巴结转移，人工智能诊断共987张存在淋巴结转移，两者诊断结果完全相同（淋巴结位置和转移数量完全相符）的图像共772张。相比于人工诊断，人工智能诊断的曲线下面积为0.886 2，单个病例的诊断时间为10 s，而影像科医师的平均判断时间为600 s。 结论： 基于深度神经网络的直肠高分辨MRI淋巴结自动识别系统具有较高的准确率，且效率高，可以辅助进行临床诊断。.