人工智能区分非典型普通型间质性肺炎及非特异性间质性肺炎的价值  

The value of artificial intelligence in distinguishing atypical usual interstitial pneumonia from nonspecific interstitial pneumonia

作　　者：黄廷端 周智鹏[1] 成戈[1] 林斌 林山月 戴文海[1] Birata Gurung HUANG Tingduan;ZHOU Zhipeng;CHENG Ge;LIN Bin;LIN Shanyue;DAI Wenhai;Birata Gurung(Department of Radiology,the Affiliated Hospital of Guilin Medical University,Guilin 541001,China)

机构地区：[1]桂林医学院附属医院放射科,广西桂林541001

出　　处：《实用放射学杂志》2024年第8期1243-1247,共5页Journal of Practical Radiology

基　　金：广西壮族自治区自然科学基金项目(2023GXNSFAA026019);桂林市科学研究与技术开发计划项目(20220139-2)。

摘　　要：目的评估人工视觉、影像组学(RAD)和深度学习(DL)对非典型普通型间质性肺炎(UIP)及非特异性间质性肺炎(NSIP)的诊断效能。方法回顾性分析300例间质性肺炎(IP)患者。经纳入和排除标准最终纳入非典型UIP 56例,NSIP 57例。所有患者在活检前均行CT检查,并随机分为训练组和测试组,开发DL和RAD 2种人工智能模型并进行训练。使用Kappa检验人工视觉评估的一致性,使用logistic回归进行受试者工作特征(ROC)曲线分析。使用DeLong检验比较各模型的曲线下面积(AUC)的差异。以AUC、准确性(ACC)为主要指标评估模型的诊断效能,并与人工视觉评估进行比较。结果2名放射科医师对非典型UIP及NSIP的CT诊断有显著的一致性(Kappa=0.852,P<0.01)。在高分辨率CT(HRCT)上,非典型UIP患者主要表现为基底部支气管扩张、蜂窝征和网状影,而NSIP患者表现为基底部网状影、磨玻璃影(GGO)及胸膜下回避征。此外,与非典型UIP患者相比,NSIP患者的GGO超出网状影以外的概率更高(P<0.05)。DL模型在非典型UIP及NSIP鉴别诊断中表现出比RAD及人工视觉更高的诊断效能(AUC:0.94 vs 0.85 vs 0.65;P<0.01)。结论CT及RAD对非典型UIP及NSIP具有一定诊断价值。基于DL模型能更准确地鉴别诊断非典型UIP及NSIP,从而进一步加强临床医师对IP的管理。Objective To evaluate the diagnostic efficacy of artificial vision,radiomics(RAD)and deep learning(DL)in distinguishing atypical usual interstitial pneumonia(UIP)from nonspecific interstitial pneumonia(NSIP).Methods A retrospective analysis was conducted on 300 patients diagnosed with interstitial pneumonia(IP).A total of 56 cases of atypical UIP and 57 cases of NSIP were included.All patients underwent CT examination before biopsy,and were randomly divided into training and test groups.DL and RAD artificial intelligence models were developed and trained.The consistency of artificial vision assessment was performed via Kappa test and receiver operating characteristic(ROC)curve analysis was performed via logistic regression.Then,the differences in the area under the curve(AUC)of different models were compared via DeLong analysis.The diagnostic efficiency of the model was evaluated via AUC and accuracy(ACC)as the main indicators and compared with the artificial vision evaluation.Results Two radiologists showed significant consistency in the CT diagnosis and interpretation of atypical UIP and NSIP(Kappa=0.852,P<0.01).On high-resolution computed tomography(HRCT),the atypical UIP patients mainly showed basal bronchiectasis,honeycombing,and reticular opacities,while,the NSIP patients showed basal reticular opacities,ground glass opacity(GGO)and subpleural avoidance sign.The probability of GGO exceeding the reticular opacities of the NSIP patients was significantly higher than that of the atypical UIP patients(P<0.05).The DL model showed significantly higher diagnostic efficiency than RAD and artificial vision in the differential diagnosis of atypical UIP and NSIP(AUC:0.94 vs 0.85 vs 0.65,P<0.01).Conclusion CT and RAD have a certain significant diagnostic value for atypical UIP and NSIP.Atypical UIP and NSIP can be more accurately differentiated and diagnosed based on the DL model,thereby,the management of IP is further strengthened in clinical practice.

关 键 词：间质性肺炎 影像组学 深度学习 

分 类 号：R563.13[医药卫生—呼吸系统] R445[医药卫生—内科学]