Smartphone image analysis OK for TB diagnosis

System seen as boon for countries where radiologists are scarce

Tuberculosis diagnosis based on cellphone camera photos of chest-rays appeared accurate enough to be of use in resource-poor areas, researchers reported at the virtual meeting of the Radiological Society of North America.

In a pilot study, x-ray images captured on smartphones equipped with an artificial intelligence app had sensitivity of 81% and specificity of 84% in detecting tuberculosis, said Po-Chih Kuo, PhD, of National Tsing Hua University, Hsinchu City, Taiwan.

"An early diagnosis of tuberculosis is crucial but is challenging for resource-poor countries with a shortage of radiologists," Kuo said. "An approach to solve this obstacle is to take photographs on chest x-ray and interpret them using smartphones. We developed a transfer deep learning-based TB detection model (TBShoNet) for chest x-ray photographs taken by a phone camera."

TBShoNet provides a method to develop an algorithm that can be deployed on phones to assist healthcare providers in areas where radiologists and high-resolution digital images are unavailable, he said. This is the first application of deep-learning technology to smartphone-captured chest x-ray photos for tuberculosis diagnosis, according to Kuo.

The AI platform "can be generalized across different datasets," he added.

For the pilot study, Kuo and colleagues obtained data from three publicly available datasets -- MIMIC-CXR, Montgomery, and Shenzhen -- which were used for model pre-training, transferring, and evaluation. A 121-layer neural network was pre-trained on the MIMIC-CXR database containing 250,044 chest radiographs with 14 pulmonary labels, which did not include tuberculosis.

The model was then re-calibrated for chest radiograph photographs by using simulation methods to augment the dataset, and then TBShoNet was built by connecting the pre-trained model to an additional two-layer neural network trained on augmented chest radiograph images in Montgomery data that included 58 tuberculosis images and 80 normal images.

Photographs were taken by five different phones of chest radiographs in Shenzhen dataset which included 336 tuberculosis images and 326 normal images.

Edith M. Marom, MD, of Chaim Sheba Medical Center/Tel Aviv University, Israel, told MedPage Today, "Having a sensitivity and specificity of about 80% is actually very good as compared to not having any interpretation at all. Such a system would work in a place where the prevalence of tuberculosis is high, where the likelihood to find other diseases which affect the lungs in similar fashion is low."

She said that the study shows that innovative thinking can fill a medical void in creative ways. "We have to think out of the box on how to maximize our efforts and improve medical care in regions which have no access to a radiologist," she said.

"In those countries in which there are no radiologists, tuberculosis is common. But it should be clarified that this method is not applicable to the entire world, was not checked on all populations, and was not compared with a final diagnosis, but only to the regular chest radiograph interpretation," said Marom, who was not involved with the study.

"One should interpret the results of this study within its limitations and context. This was a pilot study. For such an application to become commercially used, one would have to validate it, of course, on a diverse population," she noted.

Marom said many nations in Africa, for example, could benefit from a system such as the one under development by Kuo's group.

"Even when there are a few radiologists in those countries, they tend to be gathered in the main city, leaving the rest of the population with no radiologist at all," Marom said. "Thus, when a sick patient has a chest radiograph obtained to help make the diagnosis of tuberculosis, the treating medical professional is left with a chest radiograph and no interpretation. It is like not taking the film at all."