Use-Cases

Automated Chest X-Ray Reporting Using Al

Automation of chest X-ray reporting is one of the most popular applications that engineers and deep learning scientists are working on primarily due to the easy availability of publicly available datasets ChestXRayl4 by NIH (ChestX-ray8, 2019), CheXpert by Stanford University (Irvin et al., 2019) and MIMIC-CXR by Massachusetts Institute of Technology (Johnson et al., 2019). Additionally, since X-ray images are two-dimensional, they are easier to run through convolutional neural networks. Recently, there have been three peer-reviewed publications, by Singh et al. (2018), Rajpurkar et al. (2018) and Hwang et al. (2019) demonstrating technical feasibility and the potential of Al in the field of chest X-ray. We discuss three published use-cases of automated chest X-ray reporting which can reduce the workload of overburdened radiologists catering to rural India.

Normal vs. Abnormal Chest X-Ray Classification

At the European Congress of Radiology in 2019, we presented our work on retrospective validation of a high-sensitivity deep learning algorithm for normal/abnormal delineation of chest X-rays (Venugopal et al., 2019a). We evaluated the algorithm, trained on more than a million chest X-ray images, originally described by Singh et al., on a spectrum-biased dataset comprising of 430 chest X-rays with 285 abnormal and 145 normal scans. The algorithm delivered a sensitivity of >97%, with only five clinically significant false negatives (three pneumothorax, two rib fractures cases) which the algorithm was not trained to detect anyway. Since it is common knowledge that more than 50% of chest X-rays done in an outpatient setting are usually normal, such an algorithm can serve as a powerful tool not only by freeing up radiologists’ time by automatically reading normal scans but also by empowering physicians in rural India to read their own X-rays possibly through a smartphone camera. It can also be powerful in a setting of teleradiology (since the volume in such a set-up is much higher than a routine outpatient practice) enabling radiologists to focus on the abnormal cases, thereby speeding up reporting and improving quality.

Automated Diagnosis of Tuberculosis on Chest X-Ray

Putha et al., from Qure.ai, presented their work on automated tuberculosis detection using deep learning at the European Congress of Radiology in 2018 (Putha et al., 2018). Their algorithm at the time, trained on 400,000+ chest X-rays and their corresponding reports, gave an area under the curve of 0.91, 0.87 and 0.83 for the detection of tuberculosis on three external datasets. The ground truth for such data is the report and opinion of the radiologist. Given that tuberculosis is endemic in India, especially in crowded rural areas, such algorithms can go a long way in ensuring that patients’ chest X-rays are reported and can hence be acted upon in a more informed way. There are several organizations such as FIND Diagnostics, PATH and Tata Trusts which are actively working towards the complete eradication of tuberculosis - an effort which is only made possible if everyone who has tuberculosis is treated. In fact, today, there are unpublished efforts underway which aim to predict the GeneXpert status of sputum of patients, which is the gold standard for diagnosis of active pulmonary tuberculosis, from chest X-ray images using deep learning.

Al as a Quality Assessment Tool

In a presentation made at the Radiological Society of North America (RSNA) annual conference in November 2018, Sahu et al. (2018) demonstrated a use-case where a deep learning algorithm was used to parse through 3,945 randomly selected chest X-rays of adult patients scanned on an out-patient basis as part of a wellness check-up. Seven hundred and eighty-nine (20%) scans were discordant; that is, the Al-generated findings and clinical reports did not match. Subsequently, 405 of these discordant scans were re-read by three radiologists, and it turned out that in 263 (64.9%) scans, the initial clinical report was wrong. In rural India, where a primary clinical report itself is rare, such reviews performed by Al can not only help relieve the shortage of radiologists but also improve the overall quality of diagnostics and care delivery.

Automated Head CT Reporting in Emergencies

In a paper published in the Lancet, Chilamkurthy et al. describe their deep learning algorithm trained on more than 300,000 non-contrast CT scans and validated on an independent dataset from six imaging centers from across New Delhi, India (Chilamkurthy, 2018). The algorithm had an area under the curve of 0.94 for the detection of intracranial bleeds, 0.96 for the detection of skull fractures and 0.96 for the midline shift. Given that in rural India, the instantaneous availability of radiologists, and hence radiology reports, is limited, such an algorithm can play a very important role in guiding treatment by providing a provisional report with a high degree of confidence.

Chest CT Screening Using Al

In India, lung cancer constitutes 6.9% of all new' cancer cases and 9-3% of all cancer-related deaths in both sexes (Malik and Raina, 2015). Also, it is the commonest cancer and cause of cancer-related mortality in men. In rural India, smoking of indigenous cigarettes (bidi) and water-pipes

(Hookah), combined with a lack of education about its harmful effects, leads to increased risk of developing tobacco lung cancer. Another reason for the high prevalence of respiratory diseases, not necessarily cancer, is the use of wood or coal for cooking inside small homes in the villages. This makes the case for using low-dose CT for lung cancer screening in high-risk populations. Unfortunately, while multi-detector CT scanners with the capability to perform low-dose CT are generally available in rural India, radiologists with experience of picking up small nodules (<30 mm in size) are generally lacking. Aberle et al., in the National Lung Cancer Screening Trial, have demonstrated low-dose CT’s capability to reduce relative risk by 20%, but with a false positivity of up to 95% (The National Lung Screening Trial Research Team, 2011). Algorithms can help significantly reduce this false positivity. In a poster at the European Congress of Radiology in 2019, we presented a validation study of a deep learning algorithm developed by Predible Health which automatically detects and characterizes lung nodules between the size of 3and 30 mm (Venugopal et al., 2019b). The algorithm proved superior to four radiologists with 2, 5, 8 and 15 years’ experience of reading chest CT on a validation dataset of 100 CT scans with their corresponding biopsy results taken as ground truth. This work demonstrates the ability of deep learning algorithms to undertake the detection of abnormalities in a screening setting, thereby reducing the workload on radiologists working in a rural setting and improving accuracy and hence outcomes for patients.