# Performance of Computer‐Aided Detection Software in Tuberculosis Case Finding in Township Health Centers in China

## TL;DR

This study evaluates a computer-aided detection software for tuberculosis diagnosis in Chinese township health centers, finding it outperforms local radiologists in sensitivity.

## Contribution

The study demonstrates that CAD software (JF CXR-1 v3.0) significantly improves TB case detection in resource-limited settings compared to local radiologists.

## Key findings

- JF CXR-1 v3.0 showed 92.11% sensitivity for active TB diagnosis compared to 32.89% for local radiologists.
- The CAD software maintained high specificity (94.54%) similar to local radiologists (99.28%).
- CAD software could enhance TB diagnosis in primary health centers with limited resources.

## Abstract

Computer‐aided detection (CAD) software has been introduced to automatically interpret digital chest X‐rays. This study aimed to evaluate the performance of CAD software (JF CXR‐1 v3.0, which was developed by a domestic Hi‐tech enterprise) in tuberculosis (TB) case finding in China.

In 2019, we conducted an internal evaluation of the performance of JF CXR‐1 v3.0 by reading standard images annotated by a panel of experts. In 2020, using the reading results of chest X‐rays by a panel of experts as the reference standard, we conducted an on‐site prospective study to evaluate the performance of JF CXR‐1 v3.0 and local radiologists in TB case finding in 13 township health centers in Zhongmu County, Henan Province.

Internal assessment results based on 277 standard images showed that JF CXR‐1 v3.0 had a sensitivity of 85.94% (95% confidence interval [CI]: 77.42%, 94.45%) and a specificity of 74.65% (95% CI: 68.81%, 80.49%) to distinguish active TB from other imaging conditions. In the on‐site evaluation phase, images from 3705 outpatients who underwent chest X‐ray detection were read by JF CXR‐1 v3.0 and local radiologists in parallel. The imaging diagnosis of local radiologists for active TB had a sensitivity of 32.89% (95% CI: 22.33%, 43.46%) and a specificity of 99.28% (95% CI: 99.01%, 99.56%), while JF CXR‐1 v3.0 showed a significantly higher sensitivity of 92.11% (95% CI: 86.04%, 98.17%) (p < 0.05) and maintained high specificity at 94.54% (95% CI: 93.81%, 95.28%).

CAD software could play a positive role in improving the TB case finding capability of township health centers.

Flow chart of the study implementation.

CAD software might be applied to help human readers improve overall TB diagnosis in primary health center, especially in resource‐limited areas with high TB burden.Compared with local radiologists, the diagnosis of JF CXR‐1 v3.0 for active TB had a significantly higher sensitivity and a similar specificity.

CAD software might be applied to help human readers improve overall TB diagnosis in primary health center, especially in resource‐limited areas with high TB burden.

Compared with local radiologists, the diagnosis of JF CXR‐1 v3.0 for active TB had a significantly higher sensitivity and a similar specificity.