2.5D medical image classification based on NIfTI using MobileOne- LSTM

Abstract

The differential diagnosis of pulmonary tuberculosis (TB) and non-tuberculous mycobacterial (NTM) lung diseases is practically difficult due to the considerable overlap in their clinical and radiological findings, and incorrect TB diagnosis can lead to unnecessary antituberculosis treatment, drug toxicity, and delayed recognition of the true etiology. A significant portion of the current literature on TB–NTM differentiation relies on a limited number of two-dimensional (2D) CT slices or manually designed radiomic features. These approaches fail to adequately utilize the comprehensive information of the three-dimensional (3D) lung structure and often disregard probability reliability, which is critical for clinical decision support. This study presents an end-to-end deep learning approach for TB–NTM differentiation using NIfTI-format data from 3D computed tomography (CT) volumes, employing a two-stage training strategy based on 2.5D architecture (MobileOne–LSTM). Model performance was rigorously evaluated through stratified 5-fold cross-validation using accuracy, sensitivity, specificity, and F1 score; additionally, the clinical reliability of the probability outputs was analyzed using calibration curves and the Brier score. Experimental results show that the model achieved a mean sensitivity of 0.917 and an F1-score of 0.863 for TB, with an AUC of 0.830, substantially reducing the risk of false negatives at the optimal classification threshold. The satisfactory level of probability calibration demonstrates that the model outputs can be reliably used in clinical decision support scenarios. The presented study offers an original and clinically applicable approach that aims to overcome the limitations of existing methods by introducing a volumetric 2.5D deep learning and calibration-focused perspective to the TB–NTM discrimination problem.