Truncating the EfficientNet-B0 for Computer Aided Diagnosis of Tuberculosis

This study investigates the effectiveness of truncating the EfficientNet-B0 architecture for the computer-aided diagnosis of tuberculosis (TB) on chest radiograph (CXR) images. A series of truncated EfficientNet-B0 models are proposed, systematically removing blocks to reduce model complexity while maintaining diagnostic accuracy. The B0(-3) model, which eliminates three blocks, emerges as a highly efficient configuration, achieving 100% internal test accuracy on the Kaggle dataset and demonstrating robust generalization to an external Mendeley dataset. Bootstrap analysis reveals that the B0(-3) model achieves a mean accuracy of 97.38% (95% CI: 96.94%–97.84%) on the external dataset, with performance statistically overlapping that of the complete B0(-0) model (97.24%, 95% CI: 96.78%–97.69%). Despite this overlap, the B0(-3) model uses 13 times fewer parameters, making it a more efficient alternative without sacrificing accuracy. These results highlight the potential of model truncation to improve efficiency while maintaining performance, positioning B0(-3) as a promising candidate for real-world TB detection.