Differentiating Bacterial and Non-Bacterial Pneumonia on Chest CT Using Multi-Plane Features and Clinical Biomarkers.
Authors
Affiliations (5)
Affiliations (5)
- Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Kowloon, Hong Kong SAR (L.S., Y.W., M.Z., Z.L., G.R., J.C.).
- The Third People's Hospital of Longgang, Clinical Institute of Shantou University Medical College (The Third People's Hospital of Longgang District Shenzhen) /Longgang Institute of Medical Imaging, Shantou University Medical College, Guangdong, China (Y.Z.); The Seventh People's Hospital of Chongqing (Affiliated Central Hospital of Chongqing University of Technology), Chongqing, China (Y.Z.).
- Jiangxi Provincial Key Laboratory for Precision Pathology and Intelligent Diagnosis, Department of Radiology, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Jiangxi, China (L.L.); Department of Radiology, The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China, Hunan, China (L.L.).
- Department of Radiology, The First Affiliated Hospital of Hainan Medical University, Haikou, China (X.L.).
- Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Kowloon, Hong Kong SAR (L.S., Y.W., M.Z., Z.L., G.R., J.C.). Electronic address: [email protected].
Abstract
Timely and accurate classification of bacterial pneumonia (BP) is essential for guiding antibiotic therapy. However, distinguishing BP from non-bacterial pneumonia (NBP) using computed tomography (CT) is challenging due to overlapping imaging features and limited biomarker specificity, often leading to delayed or empirical treatment. This study aimed to develop and evaluate MPMT-Pneumo, a multi-plane, multi-modal deep learning model, to improve BP versus NBP differentiation. A total of 384 patients with microbiologically confirmed pneumonia (239 BP, 145 NBP) from two hospitals were included and divided into training and test sets. MPMT-Pneumo utilized a hybrid CNN-Transformer architecture to integrate features from axial, coronal, sagittal CT views and four routine inflammatory biomarkers (WBC, ANC, CRP, PCT). Poly Focal Loss addressed class imbalance during training. Performance was evaluated using Area Under the Curve (AUC), accuracy, and sensitivity on the test set. MPMT-Pneumo was benchmarked against recent deep learning models, biomarker-only models, and clinical radiologists' CT interpretations. Ablation studies assessed component contributions. MPMT-Pneumo achieved an AUC of 0.874, accuracy of 0.852, and sensitivity of 0.894 on the test set, outperforming baseline deep learning models and biomarker-only models. Sensitivity for BP detection surpassed that of less experienced radiologists and was comparable to the most experienced. Ablation studies confirmed the importance of both multi-plane imaging and biomarkers. MPMT-Pneumo provides a clinically applicable solution for BP classification and shows great potential in improving diagnostic accuracy and promoting more rational antibiotic use in clinical practice.