Multi-plane deep learning and domain-adapted radiomics for CT-based pneumonia differentiation : from algorithm development to multi-center generalization

Abstract

Background: Pneumonia is a global health challenge requiring timely and accurate diagnosis to guide effective treatment. Pathological diagnosis is invasive and slow, while computed tomography (CT)-based differentiation of bacterial pneumonia (BP) from non-bacterial pneumonia (NBP), or tuberculosis (TB) from fungal pneumonia (FP), is challenged by overlapping features and limited biomarker specificity, often leading to empirical treatment. While artificial intelligence (AI) is promising, existing models for pneumonia often focus on simpler tasks, lacking robustness across diverse clinical settings or interpretability for broader adoption. Purpose: This thesis addresses these challenges by developing and validating two novel AI models for pneumonia differentiation on chest CT: a deep learning model for BP versus NBP classification and a domain-adapted radiomics model for TB versus FP distinction. Methods and Materials: For BP vs. NBP differentiation, the MPMT-Pneumo model, a hybrid convolutional neural network (CNN)-Transformer, was developed using CT data and four inflammatory biomarkers from 384 patients across two hospitals; Poly focal loss addressed class imbalance. For TB vs. FP differentiation, a radiomics model with a novel multicenter distribution adaptation (MDA) framework was developed using CT data from 528 patients across four centers, involving automated segmentation and feature extraction from seven ROIs. Performance was evaluated by area under the curve (AUC), accuracy, and sensitivity, benchmarked against alternatives, with SHAP values for TB/FP model interpretability. Results: MPMT-Pneumo achieved AUC 0.874, accuracy 0.852, and sensitivity 0.894 for BP vs. NBP, outperforming baseline models and matching experienced radiologists' sensitivity for BP. Ablation studies confirmed MPMT-Pneumo's multi-modal design benefits. The MDA-radiomics model for TB vs. FP achieved AUCs of 0.871-0.977 in external validations, outperforming traditional classifiers. SHAP analysis offered feature insights, and t-SNE confirmed MDA's efficacy in reducing inter-center feature variability. Conclusion: This thesis developed two clinically relevant, generalizable AI solutions for CT-based pneumonia differentiation. Both the MPMT-Pneumo model and the MDA radiomics framework show significant gains in diagnostic accuracy and robustness, promising more rational antibiotic use and informed clinical decisions in diverse healthcare settings.