B99-05 CT Radiomics-Based Differentiation of Non-Tuberculous Mycobacteria From Tuberculosis Using Machine Learning

Abstract Background Differentiating non-tuberculous mycobacterial (NTM) disease from tuberculosis (TB) remains a significant clinical challenge, as both conditions present with overlapping radiological features on chest CT scans. Accurate differentiation is essential for appropriate intervention strategies. Traditional radiological interpretation relies on subjective assessment, leading to potential misclassification. We developed a radiomics-based classification model to objectively distinguish NTM from TB on CT imaging. Methods We utilized a publicly available dataset of 1,300 standard dose chest CT scans from a multicenter cohort, including culture-proven TB (n = 871) and NTM (n = 429) cases. 85 quantitative radiomic features were extracted from lung regions using PyRadiomics, including three-dimensional shape descriptors (volume, surface area), first-order intensity statistics (mean, variance, skewness), texture features derived from gray-level co-occurrence matrices, gray-level run-length matrices, and gray-level size zone matrices, as well as wavelet-transformed multi-scale features (Figure 1). The dataset was split into training (n = 909, 70%) and testing (n = 391, 30%) sets. Linear discriminant analysis (LDA) was utilized for binary classification, and performance was evaluated through accuracy, area under the receiver operating characteristic curve (AUC), and precision. Feature importance analysis identified the most discriminative radiomic features for separating NTM from TB. Results The LDA classifier achieved an AUC of 0.787 (95% CI : 0.733-0.835 )and a precision of 0.720 on the test set, outperforming the published convolutional neural network (BoTNet50) results (AUC 0.71) on the same dataset (Figure 1). Feature importance analysis revealed that intensity-based radiomic features and texture heterogeneity measures were the most discriminative parameters, with lung shape-based metrics, zone variance, large-area emphasis, inverse difference moment, inverse difference, and run percentage features, providing the strongest separation between disease entities. The high specificity of 0.912 indicates robust capability in confirming TB diagnosis, critical for avoiding unnecessary treatment delays. Conclusions CT-based radiomic features enable a lightweight interpretable machine learning model-based differentiation between NTM and TB with performance comparable to advanced deep learning approaches. The predominance of texture heterogeneity and intensity distribution features suggests that mycobacterial species create distinct quantifiable parenchymal patterns beyond visual detection. Future directions include prospective validation, integration with clinical and laboratory data, and adaptation to other lung diseases. This abstract is funded by: NIH/NHLBI K01HL163249