TL;DR
This study retrospectively evaluated a computer-aided detection (CAD) system for tuberculosis (TB) screening using chest X-rays (CXRs) in an elderly population. The CAD system significantly outperformed human radiologists in identifying suspected TB cases, suggesting its potential to improve efficiency and case-finding in high-burden settings.
What Problem Does This Paper Address?
Tuberculosis (TB) remains a significant global health challenge, with a substantial number of cases going undiagnosed annually. While chest X-ray (CXR) is a primary screening tool, its effectiveness is hampered by a shortage of trained radiologists and variability in interpretation, especially in resource-limited regions. The World Health Organization (WHO) recommends systematic screening for high-risk populations, but challenges remain in implementation and accuracy. This paper addresses the need for improved, efficient, and accurate TB screening methods, particularly for vulnerable populations like the elderly, by evaluating the performance of a specific AI-based Computer-Aided Detection (CAD) system.
Methods
The researchers conducted a retrospective cohort study using data from a 3-year active case-finding (ACF) project in Jiangshan City, China, targeting individuals aged 65 and older. They employed the JF CXR-1 v2 CAD software to analyze de-identified CXR images. This software uses convolutional neural network algorithms to identify abnormalities suggestive of TB. The study compared the performance of the CAD system against interpretations by experienced radiologists. They also analyzed the distribution and dynamic changes of CAD scores over time in relation to confirmed TB diagnoses.
Datasets
The study utilized a dataset of 49,919 chest X-ray (CXR) images from individuals aged 65 and older who participated in a 3-year active case-finding (ACF) project in Jiangshan City, China (2020-2022). Follow-up data was available for 40,741 participants in 2021 and 39,185 in 2022. Microbiologically confirmed TB cases were obtained from the Tuberculosis Information Management System (TBIMS). No new benchmark datasets were introduced; the study relied on existing clinical data.
Key Insights and Ideas
- CAD's Superiority in Case Detection: The CAD system identified a significantly higher number of suspected TB cases compared to human radiologists, suggesting it can detect subtle abnormalities that might be missed by human readers.
- Efficiency Gains with "Step" Algorithms: Combining human and AI interpretation in a "radiologist reading followed by CAD" approach significantly reduced the workload for radiologists while maintaining performance, highlighting a practical implementation strategy.
- CAD Scores as Predictive Markers: The study observed that CAD scores dynamically changed with disease progression or treatment, and higher scores (especially above the 75th percentile) were strongly associated with confirmed TB cases. This suggests CAD scores could potentially identify high-risk individuals and predict early disease development.
- Dynamic Nature of CAD Scores: The observation that CAD scores peaked around the year of diagnosis and then either increased or decreased based on treatment response offers a novel insight into disease trajectory and the potential for monitoring.
Key Results
- Prevalence: The pooled prevalence of suspected PTB reported by radiologists was 1.21%, while the CAD system reported 12.43%.
- Confirmed TB Cases: Out of 101 bacteriologically confirmed PTB cases over three years, radiologists identified 45.54% (46/101) as suspected cases, whereas the CAD system identified 83.16% (84/101).
- Workload Reduction: The "radiologist reading followed by CAD" algorithm saved at least 87% of the workload for radiologists, as they only needed to review CXRs flagged by CAD.
- High-Score Association: A majority of confirmed PTB patients (63/84) had CAD scores in the highest quartile (>75%) among those with abnormal CAD findings.
Limitations and Future Work
The authors acknowledge several limitations:
- Partial Participation: Only about half of the target population in Jiangshan City participated, potentially introducing selection bias as individuals with poorer health or less awareness might be underrepresented.
- Data Source for Confirmed Cases: Reliance on the TBIMS for confirmed cases means that unreported or misdiagnosed cases might have been missed, affecting evaluation accuracy.
- Target Population Specificity: The study focused solely on the elderly (65+). The performance of CAD and radiologists may vary across different age groups and geographical regions.
Future work should involve:
- Broader Evaluations: Conducting comprehensive evaluations across diverse regions and populations to guide future CAD applications.
- Cost-Effectiveness Analysis: Evaluating the cost-effectiveness of different CAD implementation strategies.
- Threshold Optimization: Further research is needed to determine optimal CAD score thresholds for different populations and resource settings, potentially incorporating other parameters like demographics and clinical history.
- Pilot Studies: Conducting pilot studies to test sputum collection for participants to gather unbiased data for threshold determination.
Summary updated