Imaging: cancer and tuberculosis screening
Software that reads screening mammograms and chest X-rays, scores each exam for breast cancer, lung nodules or tuberculosis and marks suspicious areas.
These tools read screening mammograms or chest X-rays, score each exam and mark suspicious areas. A 2021 BMJ systematic review of AI in breast screening (12 studies, 131,822 women) found no prospective accuracy studies, and 34 of 36 AI systems in the larger retrospective studies were less accurate than a single radiologist. A 2023 Radiology meta-analysis found standalone AI for screening digital mammography performed as well as or better than radiologists, with too few studies to judge tomosynthesis. In PRAIM, an observational study of a different product, Vara MG, in 463,094 women at 12 German sites, radiologists who chose AI support detected 6.7 cancers per 1,000, versus 5.7 in the control group, without a higher recall rate.
For tuberculosis, the World Health Organization in 2021 conditionally recommended, on low-certainty evidence, that computer-aided detection software may replace human readers of digital chest X-rays for screening and triage in people aged 15 and older.
Lunit INSIGHT CXR
Lunit INSIGHT CXR reads chest X-rays and flags 11 major thoracic abnormalities, including lung nodules, pneumothorax, consolidation and pleural effusion. The company also markets it for tuberculosis screening.
Why this grade. A pragmatic randomized trial of 10,476 health-checkup participants at Seoul National University Hospital found that radiologists using the software detected actionable lung nodules, confirmed on CT, in 0.59% of participants versus 0.25% without it.
The evidence
In the Seoul trial, run from July 2020 to December 2021 with 5,238 participants in each arm, malignant nodules were detected in 0.15% versus 0.0%, and the false-referral rate did not differ significantly (45.9% versus 56.0%). The trial was funded by a Korean Ministry of Health and Welfare grant; one author reported a Lunit research grant and another Lunit stock options.
For tuberculosis, a retrospective evaluation of five products in 23,954 people in Dhaka, Bangladesh, found Lunit INSIGHT CXR ranked third by area under the receiver operating characteristic curve (88.61%) and did not meet the World Health Organization's target product profile at 90% sensitivity; all five products outperformed radiologists. An independent evaluation of 12 products on 1,032 chest X-rays in Ho Chi Minh City found Lunit on par with an expert reader and one of three products significantly better than an intermediate reader.
- Used by
- Seoul National University Hospital's health screening center used it in a randomized trial. It was one of three products whose data informed a World Health Organization recommendation on tuberculosis screening in 2021. Lunit says its products together are used at more than 10,000 customer sites.
- FDA
- In the U.S. it is cleared only as Lunit INSIGHT CXR Triage (K211733, Nov. 10, 2021), a triage tool that prioritizes adult chest X-rays with pleural effusion or pneumothorax. Lung nodule detection, the subject of the trial behind its grade, and tuberculosis screening are not FDA-cleared indications.
- Limits
- The trial ran at one center in a low-prevalence Korean health-checkup population, and two authors reported Lunit grants or stock options. In the U.S., the FDA clearance covers only triage of pneumothorax and pleural effusion, not nodule detection or tuberculosis.
Lunit INSIGHT MMG
Lunit INSIGHT MMG analyzes 2D mammograms, provides malignancy likelihood scores and classifies lesions and highlights suspicious areas for the reading radiologist. It can also serve as an independent reader.
Why this grade. Two prospective studies in routine screening found it matched or raised cancer detection: in ScreenTrustCAD, among 55,581 women in Stockholm, replacing one of two radiologists with AI gave a 4% higher, non-inferior detection rate, and in AI-STREAM, among 24,543 women in South Korea, detection rose 13.8% without a significant change in recalls.
The evidence
ScreenTrustCAD was a prospective paired-reader study of 55,581 women aged 40 to 74. It found 261 cancers when one radiologist and AI read each exam versus 250 with two radiologists, a non-inferior result; AI alone found 246, and two radiologists plus AI found 269. In AI-STREAM, a prospective cohort in South Korea's national screening program, breast radiologists using the software detected 5.70 cancers per 1,000 women versus 5.01 without it, with recall rates of 4.53% versus 4.48%; the authors declared no competing interests.
- Used by
- Lunit says its products together are used at more than 10,000 customer sites. It was the independent AI reader in ScreenTrustCAD at Capio S:t Göran Hospital in Stockholm, AuntMinnie reported, and a reading aid in South Korea's national screening program in AI-STREAM.
- FDA
- Cleared by 510(k) on Nov. 17, 2021 (K211678). The FDA's database lists a newer version, 1.1.10, cleared April 23, 2026 (K260320), and a separately cleared tomosynthesis product, Lunit INSIGHT DBT.
- Limits
- No randomized trial has tested it. Both studies counted screen-detected cancers over short follow-up, so interval cancers, overdiagnosis and outcomes are not reported; ScreenTrustCAD ran at one site, and AI-STREAM's results are a preliminary analysis.
qXR
qXR reads chest X-rays and reports findings across the lungs, pleura, mediastinum, bones, diaphragm and heart. In tuberculosis (TB) programs, a TB score above a chosen threshold flags who should be sent for confirmatory sputum testing, such as Xpert.
Why this grade. In a multicenter implementation study that integrated qXR into routine TB case finding at Vietnam's district health facilities (24,945 chest X-rays), in which on-site physicians saw qXR's results but decided who had Xpert testing, Xpert-confirmed TB rose from 858 per 100,000 X-rays at three months to 2,112 at nine months; the study had no control group.
The evidence
In the Vietnam study, which ran from April to December 2022 with qXR 3.0, the 2,105 X-rays flagged by qXR alone were not referred for Xpert testing; agreement between physicians and the software rose from 47.3% to 76.1%, and Xpert referral from 7.8% to 9.6%. A prospective accuracy study of 2,198 symptomatic adults at Indus Hospital in Karachi, Pakistan, found the accuracy of qXR version 2 (sensitivity 0.93, specificity 0.75) non-inferior to the World Health Organization's minimum values for a triage test; sensitivity was lower in smear-negative TB.
A retrospective evaluation of five products in 23,954 people in Dhaka, Bangladesh, found qXR had the highest area under the receiver operating characteristic curve (90.81%) and was one of two products, with CAD4TB, to meet the World Health Organization's target product profile at 90% sensitivity; all five did worse in people over 60 and those with previous TB. An independent evaluation of 12 products on 1,032 chest X-rays in Ho Chi Minh City found qXR on par with an expert reader and one of three products significantly better than an intermediate reader.
- Used by
- Qure.ai says qXR is deployed at more than 3,100 sites in more than 90 countries. District health facilities in Vietnam used it in routine TB case finding, and version 2 was one of three products whose data informed a World Health Organization recommendation in 2021.
- FDA
- qXR is not FDA-cleared for TB. The FDA's records list 510(k) clearances for other modules: qXR-BT (K212690, Dec. 21, 2021), qXR-PTX-PE (K230899, Aug. 22, 2023), qXR-CTR (K231149, Sept. 22, 2023), qXR-LN for lung nodules (K231805, Dec. 22, 2023) and qXR-Detect (K251934, Jan. 16, 2026), a reading aid for adults whose indications do not mention TB.
- Limits
- No randomized trial has tested qXR's effect on TB diagnoses or outcomes, and the implementation study had no control group. Accuracy was lower in older people and those with previous TB. The World Health Organization's 2021 recommendation on computer-aided detection covers only people aged 15 and older.
Transpara
Transpara reads 2D mammograms and 3D tomosynthesis exams, sorts each exam into an elevated, intermediate or low risk category and marks suspicious regions for the radiologist. Screening programs can use the score to send low-risk exams to a single reader.
Why this grade. The MASAI randomized trial in Swedish population screening met its primary outcome, a non-inferior interval-cancer rate with Transpara-supported screening (1.55 versus 1.76 per 1,000 women); its detection and workload gains were secondary outcomes: 6.4 cancers detected per 1,000 versus 5.0 with standard double reading and a substantially lower screen-reading workload.
The evidence
MASAI randomized women to AI-supported screening, in which the Transpara score sent exams to single or double reading and marked suspicious findings, or to standard double reading. The primary analysis found interval cancers in 1.55 versus 1.76 per 1,000 women, a ratio of 0.88 (95% confidence interval, 0.65 to 1.18) that was non-inferior but not a statistically significant reduction.
In secondary outcomes, the screening-performance report found 6.4 versus 5.0 cancers detected per 1,000 (ratio 1.29), with no significant rise in recalls or false positives. An earlier safety analysis found a similar cancer detection rate with a substantially lower screen-reading workload.
- Used by
- ScreenPoint says Transpara Detection has processed more than 13 million mammograms and is in use in more than 30 countries. The MASAI trial used it at four screening sites in southwest Sweden.
- FDA
- Cleared by 510(k) as computer-assisted detection and diagnosis software for lesions suspicious for cancer. The first clearance came Nov. 21, 2018 (K181704); Transpara 2.1.0 was cleared Nov. 25, 2024 (K241831). Transpara Density, a separate product, was cleared Dec. 11, 2023 (K232096).
- Limits
- MASAI is one trial, in one Swedish region, with double reading as the comparator. Part of the extra detection was in situ cancer, and effects on overdiagnosis and breast cancer mortality were not measured.
Grades: A, randomized evidence of benefit; B, evidence from clinical use; C, accuracy studies only; D, little or no independent evidence. How the grades work. Reviews of the published evidence, not medical advice or an endorsement of any product.