EVALUATING THE ROLE OF DEEP LEARNING IN ENHANCING LUNG HEALTH ASSESSMENT: A SYSTEMATIC REVIEW AND META-ANALYSIS

Abstract

Background: With the potential to improve illness identification across many imaging modalities, deep learning (DL) models have being used more and more in pulmonary diagnostics. This meta-analysis looks at how well DL algorithms can find lung cancer, TB, and pneumoconiosis.

Methods: Six eligible studies involving DL applications in chest X-ray and histopathology were analysed. Data included AUC values, sample sizes, disease types, model architectures, and dataset sources. A Summary ROC (SROC) curve and meta-regression were used to explore performance trends and sources of heterogeneity.

Results: Good diagnostic performance was shown by the included studies' AUC values, which varied from 0.86 to 1.00.  The model by Panda et al. (2024), applied to histopathology for lung cancer detection, achieved a perfect AUC of 1.00, while Lee et al. (2020) attained an AUC of 0.99 in a large-scale X-ray-based lung cancer screening. Li et al. (2024) achieved an AUC of 0.947 with 100% sensitivity for pneumoconiosis detection. Meta-regression revealed that higher AUCs were associated with larger sample sizes, public dataset use, and specific model types such as ResNet. Disease type also contributed significantly, with lung cancer models outperforming those for tuberculosis and pneumoconiosis. A color-coded scatterplot further demonstrated the clustering of high-performing models in lung cancer with large datasets, while TB models showed moderate AUCs with smaller datasets.

Conclusion: Deep learning models, particularly those applied to lung cancer detection using chest X-ray and histopathology, exhibit high diagnostic accuracy and clinical promise. Variability in performance underscores the influence of dataset size, modality, disease type, and model architecture. To facilitate incorporation into clinical processes, future studies should concentrate on demographic diversity, interpretability, and real-world validation.