Deep learning can be useful in screening for pulmonary diseases
Computer-aided review of X-rays can help diagnose people with lung illnesses such as tuberculosis and pneumonia.
Such review, augmented by deep learning, can help lead to earlier detection and treatment of these conditions, especially in remote areas where specialists are typically in short supply, according to new research.
About one third of people in the world may be infected with tuberculosis. In 2016, there were more than 10 million cases of active tuberculosis, resulting in 1.3 million deaths. It’s the No. 1 cause of death for an infectious disease, and more than 95 percent of these deaths occurred in developing countries. Pneumonia is also prevalent, affecting 450 million people a year around the world, resulting in about 4 million deaths annually.
X-rays are crucial for screening and detecting these pulmonary diseases, but there aren’t enough radiologists to read the images, particularly in poorer nations and in rural areas. Pneumonia and tuberculosis are also difficult to diagnose because their anomalies are similar to other pulmonary diseases. Radiologists often disagree with each other on these diagnoses, according to the study authors.
The researchers, from the India-based healthcare company Artelus, are in the process of developing AI software that would use deep learning to create computer-aided screening and diagnosis approaches for these two diseases by comparing X-rays to those in its dataset, according to new research reported in arXiv.org, part of the Cornell University Library. The software has the capability of analyzing large amounts of data, and could automatically detect patterns and make predictions. The intent is to reduce the burden on radiologists and enable screening on a larger scale.
“Efforts to eliminate the TB epidemic are challenged by the persistent social inequalities in health, the small number of local healthcare professionals, and the weak healthcare infrastructure found in resource-poor settings,” the study authors say.
The computer system being designed would be user-centered and deployable on a mobile chip.
The researchers have already created an automatic artificial intelligence-aided screening model using a deep network with 45 million trainable parameters and 211 layers. The software was trained using four publicly available datasets of frontal chest X-ray images consisting of more than 89,000 cases. The researchers also used data augmentation, such as flipping the X-ray images and changing the pixel values randomly, using hue and contrast, to expand the database further.
The model surpassed the level of radiologist performance with 96 percent sensitivity and 91 percent specificity for pneumonia, and 92.5 percent sensitivity and 85 percent specificity for tuberculosis.
The researchers now aim to incorporate radiographs of lateral views into the model and integrate medical histories and lifestyle information to enhance the model’s predicting capabilities.
“The ability to do faster more accurate diagnosis leads us to our primary goal of reducing the time it takes to screen population in the vulnerable areas or poor socio-economic zones for early diagnosis of TB and to minimize the spreading of the germs and transmitting and infecting others in the group. The ultimate goal of our research is to reduce patient wait times for being diagnosed with this infectious disease by developing new machine learning and mobile health techniques to the TB screening and diagnosis problem,” the researchers say.