利用深度学习实现CT图像上肾脏肿瘤径线自动测量的临床可行性

Clinical feasibility study of automatic measurement of renal tumor diameter on CT images using deep learning

目的:评估U-Net分割模型对肾脏肿瘤分割及径线测量的准确性。方法:回顾性收集本院PACS中2019年5月-2019年11月经手术病理证实的肾肿瘤患者的影像图像及结构式报告。排除未行手术治疗或病理结果未知的病例、图像及报告质量不合格及既往手术史的病例后,共纳入154例数据。从肾脏肿瘤结构式报告中导出医生测量值。利用U-Net模型自动分割肾脏肿瘤,并采用最小体积包围盒算法得到模型测量值。两位影像医生标注肾肿瘤,并采用最小体积包围盒算法得到参考值。对参考值、医生测量值、模型测量值三组数据进行统计学分析。结果:模型测量值与参考值相比,肾肿瘤短径、中径、长径之间的差异无统计学意义(P>0.05)。肿瘤各径线的医生测量值均小于参考值,差异均具有统计学意义(P<0.05)。肿瘤各径线的模型测量值大于医生测量值,差异均具有统计学意义(P<0.05)。医生测量值与模型测量值的一致性高。结论:基于U-Net的肾肿瘤分割模型自动测量肿瘤径线具有临床可行性。

Objective:To evaluate the accuracy of the U-Net model for renal tumor segmentation and diameter measurement.Methods:The images and structured reports of patients with renal tumors confirmed by surgery and pathology in PACS of our hospital from May 2019 to November 2019 were retrospectively collected.A total of 154 patients were included after excluding cases without surgical treatment or unknown pathological results, substandard images and reports, and previous surgical history.Radiologists’ measurements were derived from the Kidney Tumor Structural Report.Renal tumors were automatically segmented using a U-Net model, and the model measurements were obtained using the minimum volume bounding box algorithm.Two radiologists annotated the renal tumors and used the minimum volume bounding box algorithm to obtain the reference values.Statistical analysis was performed on the three groups of the reference value, radiologists’ measurement value, and model measurement value.Results:Comparing the measured values of the model with the reference values, there was no statistically significant difference in the short diameter, medium diameter, and long diameter of renal tumors(P>0.05).The radiologists’ measured values were lower than the reference values in tumor diameter, and the difference was statistically significant(P<0.05).The model measurement value of each tumor diameter was greater than the radiologists’ measurement value, and the difference was statistically significant(P<0.05).There was a good agreement between the radiologists’ measurement and the model measurement.Conclusion:It is clinically feasible to automatically measure the renal tumor’s diameter based on the U-net model.