Fusion Mamba-driven 3D End-to-end Model for Diagnosis and Prognostic Prediction of Pneumonia Using CT Scans

Pneumonia is a prevalent and potentially fatal respiratory disease, and how to accurately diagnose and predict the in-hospital prognosis of pneumonia patients based on limited information has always been a critical clinical challenge. This work aims to develop a priori-guided collaborative model that fully utilizes the detailed information contained within three-dimensional (3D) computed tomography (CT) data to improve the accuracy and efficiency of both initial diagnosis and prognostic prediction for pneumonia patients. The proposed model is composed of stacked Multi-scale Collaborative Modules (MCMs), each of which contains a self-designed triple-direction Mamba (TD M) branch, a 3D convolutional (3D-Conv) branch, and a Fusion Mamba (FM) block. Within each MCM, the TD-M branch extracts long-range global dependencies from 3D CT data, while the 3D-Conv branch complements local spatial and morphological representations at the same scale. The FM block further integrates these complementary features to enhance feature interaction within each scale. After the stacked MCMs, an MLP head consisting of pooling layers and fully connected layers is employed to achieve pneumonia diagnosis and prognostic prediction. A total of 1677 CT scans were collected to train, validate, and test the proposed model in two clinical tasks. Experimental results demonstrate that the proposed model achieves superior performance in pneumonia diagnosis and prognostic prediction, with accuracies of 92.37±1.56% and 86.75±1.16% in validation set, and 91.33±0.50% and 86.60±1.60% in testing set, respectively. Ablation and comparison experiments further prove the feasibility of the proposed model, indicating its considerable potential for clinical application.