Abstracts

RETURN TO ABSTRACT LISTING


E1763. Automated Detection of Tuberculosis in Radiographs for Population Based Screening Programs
Authors
  1. Amit Kharat; DeepTek Imaging Private Limited
  2. Gunjan Naik; DeepTek Imaging Private Limited
  3. Priyam Choudhury; DeepTek Imaging Private Limited
  4. Rohit Lokwani; DeepTek Imaging Private Limited
  5. Swaraj Kasliwal; DeepTek Imaging Private Limited
  6. Sudeep Kondal; DeepTek Imaging Private Limited
  7. Tanveer Gupte; DeepTek Imaging Private Limited
Objective:
The tuberculosis (TB) control program of India is one of the largest public health programs in the world (1). In spite of remarkable success it still faces a lot of challenges. Hence, a deep learning model for chest X-rays was developed and deployed at a live TB screening site to automatically screen the population for the presence or absence of TB. This is especially important in India, where the burden of pulmonary Tuberculosis is extremely high and radiologist to patient ratio is extremely low (2). In this study, we discuss our observations regarding the performance of the model, and challenges faced by the model while working as a screening tool in population based screening, where prevalence of a disease may be lower than hospital based screening where prevalence of diseases are typically higher.

Materials and Methods:
From a TB population screening program run by a government body in collaboration with a public health foundation, we obtained a CXR dataset consisting of 44,623 images taken in mobile diagnostic vans during the period April 2019 to September 2019. These images were anonymized and annotated by a qualified radiologist who marked each image as either TB-positive or TB-negative. An image that contained indications of any of the following pathologies was marked as TB-positive: nodular shadows, infiltrates, pleural effusion, pneumonia or consolidation-like features, fibrosis, pleural thickening, granuloma, bronchiectasis, scarring, lymph node, and calcified pleural plaques (3). We trained our convolution neural network with 7,039,554 parameters on these images. During live testing, the model was used to generate predictions for 18,170 CXRs taken from October 2019 to December 2019.

Results:
We observed the following performance during live deployment: accuracy = 69%, sensitivity = 90%, specificity = 68%, and AUROC = 0.91.

Conclusion:
Recognizing that early diagnosis is the key to eradicating tuberculosis, governments around the world have started active case-finding through large population screening programs. In rural and remote areas, such programs can benefit greatly from neural network-based models arming edge devices. Mass pre-screening using neural network models on digital CXRs can provide an affordable and convenient population screening solution.