1. Introduction
Pneumonia remains a leading cause of childhood mortality in regions with limited access to radiologists, and automated triage of chest radiographs can help prioritise cases for urgent specialist review where reporting turnaround is otherwise measured in days.
2. Methodology
A MobileViT-XS architecture, combining lightweight convolutional stages with transformer blocks, was fine-tuned on 12,800 frontal chest radiographs drawn from combined paediatric and adult public datasets, labelled as pneumonia-positive or normal, with the classification threshold selected via a precision-recall curve to prioritise sensitivity given the screening use case.
3. Results
At the selected operating threshold, the model achieved 95.6 percent sensitivity and 91.3 percent specificity on a held-out test set, with an AUC of 0.96. Inference averaged 41 milliseconds per image on a standard laptop CPU without GPU acceleration, and the full model occupied 9.7MB, supporting deployment on modest clinic hardware.
4. Conclusion
A compact vision transformer architecture can deliver radiologist-comparable screening sensitivity for pneumonia while running efficiently on non-specialist hardware. Future work will evaluate performance across a wider range of X-ray equipment vendors to assess generalisation.
References
[1] Mehta S. and Rastegari M., MobileViT: Light-weight, general-purpose, and mobile-friendly vision transformer, ICLR, 2022. [2] Rajpurkar P. et al., CheXNet: Radiologist-level pneumonia detection, arXiv, 2017.