探讨基于卷积神经网络对颅底骨折CT图像精准诊断的应用价值

The Application Value of Accurate Diagnosis of CT Image of Skull Base Fractures based on Convolutional Neural Network

目的:探讨卷积神经网络(CNN)在颅底骨折CT诊断的应用价值。方法:回顾性搜集3 100例颅底骨折患者及2 467例正常患者的颅骨CT图像数据,经纳排标准筛选,最终选用2 488例颅底骨折及1 628例正常患者的颅底CT图像数据。对CT图像进行骨折标注后,随机分配训练集和测试集后。通过CNN构建颅骨区域识别算法模型和颅骨骨折检测算法模型,随后在测试中以颅底骨折区域识别和头颅骨折、颅底骨折对模型进行验证,验证指标为精准率(precision)、召回率(recall)、平均诊断耗时;与人工组(低年资放射科医师)测试进行诊断效能对比。结果:通过CNN运算获得的稳定模型后进行测试对比,结果显示全颅底区域骨折、前、中、后颅底骨折精准度均<0.5,低于人工组(均> 0.63);召回率> 0.89,均优于人工组(均<0.8);平均诊断时间为(3.12±2.67)s,明显少于人工组诊断时间。分别在颅底骨折区域测试中,精准度率:前颅底>中颅底>后颅底,召回率:中颅底>后颅底>前颅底。结论:基于CNN颅底骨折算法模型对于颅脑外伤患者CT诊断颅底骨折在召回率、诊断耗时均优于人工测试结果,在辅助临床诊断、降低漏诊及诊断耗时方面具有一定的价值。

Objective: To explore the application value of convolutional neural network(CNN) in CT diagnosis of skull base fractures. Methods: The skull CT image data of 3100 patients with skull base fractures and 2 467 normal patients was collected retrospectively. After the standard nanofiltration and actual model calculation, the skull base CT image data of 2 488 patients with skull base fractures and 1 628 normal patients were selected. The CT images were labeled and randomly assigned into training set and test set. The skull area discrimination algorithm model and skull base fractures detection algorithm model were established by CNN, then we performed verification on the models through skull base area discrimination, skull fractures and skull base fractures in the test. The detection indexes included precision, recall and average diagnosis time consumption. We carried out comparisons of diagnostic efficacy with the artificial group(junior radiologist) test. Results: We carried out test comparisons on the steady models obtained by CNN algorithm, the results showed that the accuracy of the whole skull base fractures(including the anterior, middle and posterior skull base fractures) was less than 0.5, which was lower than that of the artificial group(all higher than 0.63);The recall rate > 0.89 was better than that of the artificial group(all < 0.8);The average diagnosis time was(3.12 ±67)s, significantly less than that of artificial group. In the area test of skull base fractures, the accuracy rate was anterior skull base > middle skull base > posterior skull base while the recall rate was middle skull base > posterior skull base > anterior skull base.Conclusion: The algorithm model of skull base fractures based on CNN is superior to the artificial test results in recall rate and diagnosis time consumption for CT diagnosis of skull base fractures in patients with craniocerebral trauma, which has certain value in assisting clinical diagnosis, reducing missed diagnosis and diagnosis time consumption.