The diagnostic model from semi-supervised cross modality transformation improved the distinguished ability of X-rays...

Early diagnosis of tuberculosis is particularly difficult in resource-poor areas. Traditional chest X-rays (CXR) have limited accuracy, while CT scans are costly and involve radiation exposure. The study aims to improve the diagnostic accuracy of routine X-rays for pulmonary tuberculosis to approximate the performance of CT scans through building Artificial Intelligence (AI) model, suitable for primary healthcare settings lacking CT facilities.

In this study, datasets from our hospital and two open-source datasets, namely the Shenzhen Hospital dataset (CHNCXR) and the Montgomery County dataset (MC), were included. A semi-supervised cross-modality transformation computational model was employed to independently train deep learning models based on X-ray and CT images. Transfer learning was utilized for pre-training on ImageNet, and the model performance was evaluated using 5-fold cross-validation.

In the evaluated patients, MX'(final augmented X-ray model) shows a standout performance in diagnosing pulmonary tuberculosis (PTB) using chest X-rays, with a 6% increase in high precision and a 1.8% increase in specificity, significantly surpassing the original X-ray model MX(X-ray model). Although MX' has a lower sensitivity (0.778) compared to MX (0.815), its overall balance makes it highly suitable for initial screenings. The model's ability to prioritize accuracy and specificity highlights its potential for effective deployment in clinical scenarios with follow-up testing options.

The novel diagnosis model based on the AI method strikes a meaningful balance between precision and accessibility. This makes MX' a practical alternative in resource-limited settings, offering a more efficient and scalable solution for tuberculosis diagnosis and screening.