Computed tomography-based radiomics prediction model for differentiating invasive pulmonary aspergillosis and pneumocystis jirovecii pneumonia

Zhiguo Peng¹Xingzhe Gao²Miao He³Dong XinYue⁴Dongdong Wang⁵Zhengjun Dai⁶Dexin Yu⁵Huaibin Sun¹Jun Tian¹Yu Hu^7*

• ¹Department of Organ Transplantation, Qilu Hospital, Cheeloo College of Medicine, Jinan, China
• ²Department of Anesthesiology, Qilu Hospital (Qingdao), Cheeloo College of Medicine, Shandong University, Qingdao, Qingdao, China
• ³Department of Medical Oncology, Qilu Hospital (Qingdao), Cheeloo College of Medicine, Shandong University, Qingdao, China
• ⁴Department of Oncology, Qilu Hospital of Shandong University, Dezhou Hospital, Dezhou, PR China., Dezhou, China
• ⁵Department of Radiology, Qilu Hospital, Cheeloo College of Medicine, Shandong University, Jinan, China
• ⁶Scientific Research Department, Huiying Medical Technology Co., Ltd, Beijing, PR China, Beijing, China
• ⁷Department of Medical Oncology, Qilu Hospital, Shandong University, Jinan, China

Background: Pneumocystis jirovecii and Aspergillus fumigatus are important pathogens that cause fungal pulmonary infections. Because the manifestations of pneumocystis jirovecii pneumonia (PJP) or invasive pulmonary aspergillosis (IPA) are difficult to differentiate on computed tomography (CT) images and the treatment of the two diseases is different, correct imaging diagnosis is highly significant. The present study developed and validated the diagnostic performance of a CT-based radiomics prediction model for differentiating IPA from PJP.Methods: Ninety-seven patients, 51 with IPA and 46 with PJP, were included in this study. Each patient underwent a non-enhanced chest CT examination. All patients were randomly divided into two cohorts, training and validation, at a ratio of 7:3 using random seeds automatically generated using the Radcloud platform. Image segmentation, feature extraction, and radiomic feature selection were performed on the Radcloud platform. The regions of interest (ROI) were manually segmented, including the consolidation area with the surrounding ground-glass opacities(GGO) area and the consolidation area alone. Six supervised-learning classifiers were used to develop a CT-based radiomics prediction model, which was estimated using the receiver operating characteristic (ROC) curve, area under the curve (AUC), sensitivity, specificity, precision, and f1-score. The radiomics score was also calculated to compare the prediction performance.Results: Classifiers trained with the consolidation area and surrounding GGO area as the ROI showed better prediction efficacy than classifiers trained using only the consolidation area as the ROI. The XGBoost model performed better than the other classifiers and radiomics scores in the validation cohort, with an AUC of 0.808 (95%CI, 0.655-0.961).Conclusions:This radiomics model can effectively assist in the differential diagnosis of PJP and IPA. The consolidation area with surrounding GGO was more suitable for ROI segmentation.