B41-29 Diagnostic Accuracy of Deep Learning Models for Interstitial Lung Disease on Chest CT: A Systematic Review and Meta-Analysis

Abstract Background Interstitial lung diseases (ILDs) include a broad range of parenchymal disorders, where computed tomography (HRCT) plays a vital diagnostic role. However, overlapping imaging features and interobserver variability often hinder accurate interpretation. Advances in artificial intelligence (AI) enhanced CT offers improved detection of subtle imaging patterns. While AI has been studied in thoracic imaging, its diagnostic performance in ILDs remains underexplored. Synthesizing current evidence is essential to assess its clinical reliability and applicability. Methods This Systematic Review and Meta Analysis followed the PRISMA guidelines and has been registered on PROSPERO under the registration number CRD420251184238. A comprehensive literature review was conducted on PubMed and Embase from inception to October 2025 for studies applying AI to detect ILD using CT scan. Eligible studies reported diagnostic performance using sensitivity, specificity and accuracy. Random-effect models in “R” calculated pooled sensitivity, specificity and accuracy. Results Pooling 34 deep learning studies revealed that AI algorithms maintain diagnostic potential for ILD across diverse patient populations, imaging platforms, and model designs. Using the random-effect model, the pooled accuracy was 0.894 (95% CI 0.829-0.937), confirming that deep learning models deliver consistent ILD classification accuracy despite diverse datasets, model architectures, and imaging acquisition differences. Sensitivity synthesis across 40 studies showed that deep learning identified ILD, with a pooled sensitivity of 0.890 (95% CI 0.842-0.926), demonstrating the model’s high reliability and minimizing missed diagnoses. Specificity synthesis across 31 studies indicated the model’s excellent ability to distinguish between ILD and non-ILD patients, with a pooled specificity of 0.922 (95% CI 0.861-0.957), and multiple external validation datasets reached 1.00 specificity, confirming meticulous exclusion of non-ILD patients. Although heterogeneity was high, it reflected highly variable dataset and ILD subtype distribution, rather than inconsistency in model performance, reinforcing robustness of deep learning models across varied clinical environments. Conclusion Deep learning, when applied to CT, demonstrated high accuracy, sensitivity, and specificity for diagnosis and classifying interstitial lung diseases, with performance that approached expert interpretation. Diagnostic credibility remained stable across diverse datasets, architectures, and clinical sites, showing that deep learning remains reliable even when applied to real-world, heterogeneous imaging environments. By improving the consistency and promptness of ILD recognition, deep learning reduces ambiguity, facilitates diagnosis, and supports earlier treatment. Employing deep learning in ILD evaluation enables faster diagnostic accuracy, reduces interpretation variability, and supports more informed clinical decisions. This abstract is funded by: None