The diagnostic classification value of an AI-based CT radiomics model for bacterial pneumonia
基于人工智能的CT放射组学模型对细菌性肺炎的分类诊断价值

摘要

目的 建立CT影像特征分析模型，鉴别肺炎链球菌、金黄色葡萄球菌、肺炎克雷伯菌诱发的肺部炎症，方便临床尽早判定致病菌种，进而调整适配的治疗方案。方法 筛选2023年1月至2024年12月院内确诊的80例细菌性肺炎病例，依据病原学检测结果划分组别，肺炎链球菌感染28例，金黄色葡萄球菌感染25例，肺炎克雷伯菌感染27例。全部病例均完成胸部CT平扫检查，提取病变部位126项影像数据，通过组内相关系数核验数据可靠性，保留合格指标后运用最小绝对收缩与选择算子（LASSO）算法筛选关键参数，构建多分类诊断模型。把病原培养、核酸检测结果当作评判标准，评定模型诊断水平，同时对比两位经验丰富影像医师的诊断结论。结果 该模型对三类致病菌肺炎的诊断准确率分别为89.3%、88.0%、92.6%，整体诊断准确率达到90.0%。ROC曲线统计得出，三组病例对应的曲线下面积依次为0.912、0.905、0.931，模型具备出色的鉴别诊断实力。结论 借助CT影像数据搭建的分析模型，可以精准分辨不同病菌造成的肺部影像差别，诊断水平高于常规人工阅片，能够辅助临床快速判定感染菌种，具备较高的实际临床使用价值。

Abstract

Objective To establish a CT imaging feature analysis model for differentiating pulmonary inflammation caused by Streptococcus pneumoniae, Staphylococcus aureus, and Klebsiella pneumoniae, enabling early clinical identification of pathogenic bacteria and subsequent adjustment of appropriate treatment regimens.
Methods A total of 80 cases of bacterial pneumonia diagnosed in the hospital from January 2023 to December 2024 were selected and categorized based on etiological testing results: 28 cases of Streptococcus pneumoniae infection, 25 cases of Staphylococcus aureus infection, and 27 cases of Klebsiella pneumoniae infection. All cases underwent plain chest CT scans, with 126 imaging parameters extracted from the affected areas. Data reliability was verified using intra-group correlation coefficients, and qualified parameters were retained for key parameter selection via the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to construct a multi-classification diagnostic model. Pathogen culture and nucleic acid test results served as evaluation criteria for assessing model performance, with comparative analysis conducted between diagnoses by two experienced radiologists.
Results The model demonstrated diagnostic accuracy rates of 89.3%,88.0%, and 92.6% for infections caused by the three pathogens, respectively, achieving an overall diagnostic accuracy of 90.0%. ROC curve statistics showed area under the curve values of 0.912,0.905, and 0.931 for the three groups, confirming the model’s superior discriminative diagnostic capability.
Conclusion   The analytical model developed using CT imaging data accurately distinguishes pulmonary imaging differences caused by different pathogens, outperforming conventional manual image interpretation in diagnostic accuracy and offering significant clinical utility for rapid identification of infectious pathogens.