Computer-aided Detection Finds Tuberculosis in Underserved Regions

Artificial intelligence-trained computer networks have the potential to accurately diagnose tuberculosis (TB). A retrospective study published in Radiology provides a potential model for cost-effective evaluation of chest radiographs in medically underserved regions where TB is prevalent.

Radiologists Paras Lakhani, MD, and Baskaran Sundaram, MD, from Thomas Jefferson University Hospital in Philadelphia, Pennsylvania, sought to examine the efficacy of computer networks in diagnosing TB based on lung images.

The study of 1007 posteroanterior chest radiographs (492 TB-positive cases, 515 healthy controls) comprised 4 datasets from the United States, China, and Belarus and 2 prominent deep convolutional neural networks (DCNN) — well-trained computer systems that rely on layers of data for decision making.

Two DCNNs, AlexNet and GoogLeNet, were used to evaluate lung images. Both networks employed trained and untrained models to determine whether the lung was positive for TB. The networks’ diagnostic abilities were evaluated separately and together.

The most accurate method for diagnosing TB was an ensemble of the 2 networks working in concert (area under the curve [AUC]: 0.99; 95% CI, 0.96-1.00). The sensitivity of the ensemble was 97.3% and the specificity was 94.7%. AlexNet and GoogLeNet networks’ trained models demonstrated greater diagnostic accuracy than their untrained models (AUC 0.98 vs 0.90 and 0.97 vs 0.88, respectively, P <.001).

In the 13 instances in which neither DCNN could distinguish TB, a cardiothoracic radiologist correctly diagnosed those cases (97.3% sensitivity and 100% specificity), suggesting that a second evaluation by a radiologist could further improve highly accurate diagnoses.

“This technology will need to be proven more carefully in a clinical environment, and also evaluated on subtle disease that is more difficult to detect,” said Dr Lakhani in an email interview with Infectious Disease Advisor. “Even though the results were very good in the paper, there is still room for improvement with regard to the model, and training on a larger volume of cases would be helpful for future directions.”

Study Limitations

• The DCNNs were trained to diagnose TB only and therefore could not detect other abnormalities that a radiologist would be able to discern
• If the same algorithms were applied to non-TB endemic regions, false-positives could result
• The retrospective study used images that may have had greater resolution, thus improving diagnostic accuracy