基于深度学习与功能磁共振的人工智能前列腺癌诊断效能  

Efficiency Assessment of Prostate Cancer Diagnosis Based on Artificial Intelligence Based on Deep Learning and Functional Magnetic Resonance Imaging

作　　者：李凌昊 胡怡音 孟广明 黄子丹 李致勋[3] 姜建[2] LI Linghao;HU Yiyin;MENG Guangming;HUANG Zidan;LI Zhixun;JIANG Jian(Department of Imaging,the First Affiliated Hospital of Nanchang University,Nanchang 330006,China;不详)

机构地区：[1]南昌大学第一临床医学院,江西南昌330027 [2]南昌大学第一附属医院影像科,江西南昌330006 [3]南昌大学信息工程学院,江西南昌330031

出　　处：《中国医学影像学杂志》2021年第4期385-389,共5页Chinese Journal of Medical Imaging

基　　金：国家自然科学基金(81960313);国家自然科学基金(81960325)。

摘　　要：目的评估基于深度学习网络的人工智能通过扩散加权成像(DWI)与表观扩散系数(ADC)对前列腺癌的诊断价值。资料与方法抽取南昌大学第一附属医院数据库中进行过前列腺MRI扫描并取得病理结果的112例患者,其中前列腺癌52例,由3名影像医师对图片进行标记后,导入人工智能系统进行训练,使用未标记数据进行分割准确性测试,并利用ADC图像对残差网络(ResNet)ADC区域分割的前列腺癌诊断能力进行测试。使用交并比指标计算分割准确率。使用受试者工作特征曲线及曲线下面积(AUC)评估定性诊断模型效能。结果卷积神经网络的ADC与DWI准确度分别为61.34%、57.35%;ResNet的ADC与DWI准确度分别为60.05%、63.08%;ResNet定性诊断模型的AUC为0.782,准确度、敏感度和特异度分别为69.39%、54.50%、73.68%。结论深度学习网络在前列腺病灶分割和定性分析上均显示出较大的优势,具有一定的临床实用性。Purpose To evaluate the diagnostic value of artificial intelligence for prostate cancer based on deep learning network by diffusion weighted imaging(DWI)and apparent dispersion coefficient(ADC).Materials and Methods In total,112 patients underwent MRI and pathological examination were extracted from the database of the first affiliated hospital of Nanchang university.Of the all patients,52 patients were diagnosed as prostate cancer.Imaging data were labeled by three radiologists and imported artificial intelligence system for training.The untagged data were used for segmentation accuracy test,and residual network(ResNet)was tested for the ADC region segmentation of diagnostic ability in prostate cancer.Intersection over union(IOU)was used to calculate the segmentation accuracy.Receiver operating characteristic curve and area under curve(AUC)were used to evaluate the effectiveness of the qualitative diagnostic model.Results The accuracy of ADC and DWI of convolutional neural network was 61.34%and 57.35%,respectively.The accuracy of ADC and DWI of ResNet was 60.05%and 63.08%,respectively.The AUC of ResNet of qualitative diagnosis model was 0.782;the accuracy,sensitivity and specificity was 69.39%,54.50%and 73.68%,respectively.Conclusion Deep learning network shows great advantages in the segmentation and qualitative analysis of prostate lesions,and has certain clinical practicability.

关 键 词：前列腺肿瘤 磁共振成像 扩散加权成像 深度学习 表观扩散系数 卷积神经网络 残差网络 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.25[医药卫生—诊断学]