Can AI chest X-ray software be used for TB screening in a district programme?
What the WHO guideline actually says
The governing document is the WHO Consolidated Guidelines on Tuberculosis: Module 2 — Screening: Systematic screening for tuberculosis disease, published 22 March 2021 (WHO/2021/9240022676). Its recommendation reads: "Computer-aided detection (CAD) software programmes may be used in place of human readers for interpreting digital chest X-rays for screening and triage for TB disease."
Three qualifiers carry most of the operational weight. First, "may be used" — this is a conditional recommendation with low certainty of evidence, meaning the benefits probably outweigh the harms but the evidence base is weak and the right choice depends on context. It is not a strong recommendation and not an endorsement of any named product. Second, "digital chest X-rays" — the recommendation is scoped to digital radiography. Third, "screening and triage" — CAD sits at the screening step of the pathway, ahead of diagnosis, not at the diagnostic step.
There is also an explicit age boundary. WHO states that it "does not yet recommend the use of CAD to screen children and adolescents younger than 15." A district campaign using CAD therefore needs a separate, non-CAD route for people under 15, applied before triage rather than after.
Who WHO says to screen in the first place
A software decision sits downstream of a targeting decision. Module 2 sets out which groups systematic screening should prioritise, and the strength of those recommendations varies.
| Recommendation strength | Groups named in WHO Module 2 (2021) |
|---|---|
| Strong (benefits clearly outweigh harms) | People living with HIV, at each health facility visit; household and close contacts of TB patients; incarcerated populations, at entry, annually and on release; miners and workers exposed to silica dust. |
| Conditional (benefits probably outweigh harms, with trade-offs) | Structural risk groups including the urban poor, homeless people, migrants, refugees, internally displaced persons and indigenous populations (very low certainty); the general population where TB prevalence is 0.5% or higher; health facility attendees with clinical risk factors in high-prevalence settings of 100 per 100,000 or higher. |
This matters because, even at an identical threshold, the share of flagged people who turn out to have bacteriologically confirmed TB differs between a household-contact cohort and a walk-in general population.
What CAD replaces, and what it does not
CAD reads a digital chest radiograph and returns an abnormality score. In the WHO framing it substitutes for the human reader at the interpretation step, and for nothing else in the chain.
- Image acquisition. A functioning digital X-ray unit, a trained radiographer, correct positioning and acceptable image quality remain prerequisites. Poor acquisition is not fixed by software downstream.
- Clinical judgement. A flag is one input. Symptom history, HIV status, contact history and prior TB treatment still shape what happens next.
- Bacteriological confirmation. A screen-positive result is an instruction to test, not a conclusion. Nothing in the screening step establishes a diagnosis.
- Record-keeping and follow-up. A score on a workstation is not a programme record until someone enters it into the district's tracking system and acts on it.
Read as a workflow substitution rather than a clinical one, the question is narrow: can this interpretation step run without a human reader here, at a locally calibrated threshold, with confirmatory capacity sized to the flags it will generate?
Why a CAD score is a threshold decision, not a yes or no
CAD output is not "TB" or "no TB." It is a continuous abnormality score, and a cutoff applied to that score determines who is referred onward. WHO deliberately does not fix one universal global cutoff. It has published a calibration toolkit — "Determining the local calibration of computer-assisted detection (CAD) thresholds and other parameters" — with an operational research protocol and a web-based analysis tool, so implementers can determine the appropriate threshold for a specific CAD implementation. The stated reason is blunt: WHO identified substantial variation in the diagnostic accuracy, in both sensitivity and specificity, of CAD programmes across settings.
A vendor default threshold is therefore a starting hypothesis, not a settled configuration. Calibration is a local trade-off exercise:
- A lower threshold flags more people. Sensitivity rises, more confirmatory tests are consumed, and pressure moves onto sputum collection, transport and molecular testing throughput.
- A higher threshold flags fewer people. Specificity rises and testing load falls, but more people who have TB pass through the screen unflagged.
- The right point on that curve is set by confirmatory testing capacity, the population screened, cost per test, and how the programme weighs a missed case against an extra test.
Because that balance is local, threshold setting belongs in an operational research protocol with a review point — not a setting nobody revisits after installation.
A flagged X-ray still requires confirmatory testing
Under WHO's testing framework, confirmation runs on a molecular WHO-recommended rapid diagnostic test (mWRD) applied to a sputum sample. The named tests are Xpert MTB/RIF and Xpert MTB/RIF Ultra, cartridge-based real-time PCR on the GeneXpert platform returning results in under two hours and also assessing rifampicin resistance, and Truenat MTB and Truenat MTB Plus, chip-based real-time micro-PCR returning results in roughly an hour.
Scope note: no single WHO document states a rigid rule of the form "if CAD is positive, perform Xpert MTB/RIF." Both sit within the same screening-to-diagnosis architecture, so the link between a positive screen and a named confirmatory test is a programme design decision, not a quoted WHO instruction.
Where this lands inside India's district TB structure
In India, TB services run under the National Tuberculosis Elimination Programme (NTEP), implemented within the National Health Mission with resources shared between state and central governments. The District TB Centre (DTC) is described as the nodal point for TB control activities in the district, covering diagnosis, treatment of both drug-susceptible and drug-resistant TB, and follow-up. The District TB Officer holds overall responsibility for managing NTEP at district level as per programme guidelines and the guidance of the District Health Society, covering planning, training, supervision and monitoring.
The cascade runs in four linked stages: screening and case finding, through passive, active and intensified case finding; diagnosis, on a tiered laboratory network of microscopy, NAAT and culture-based testing provided free of cost; treatment initiation, with prompt start within three days of diagnosis and fixed-dose drug combinations; and follow-up, including contact tracing, adherence monitoring and linkage to social welfare schemes.
CAD touches one of those four stages. It can change how the screening step is staffed and how quickly an interpretation is available. It cannot change whether the handoff from a flag to a sputum sample to a treatment card actually happens — and that handoff is where screening campaigns lose people.
The questions to settle before adoption
| Question | Why it decides the outcome |
|---|---|
| Who operates the X-ray unit, and who is trained on what the score means? | CAD replaces a reader, not a radiographer. Image quality stays a human responsibility. |
| Does CAD run on-device or off-site, and what connectivity exists? | Decides whether a score arrives while the person is still present or after they leave. |
| Where is the score written down? | A score living only on a machine is not a record and cannot be audited or followed up. |
| What threshold is in use, set by whom, on what local data? | An unexamined default silently sets the district's sensitivity and testing load. |
| Can confirmatory testing absorb the flag volume? | Flags beyond mWRD capacity become backlogs, and backlogs become people lost before confirmation. |
| Who owns the screen-positive and test-pending list? | Without a named owner, results are generated that nobody closes out. |
| What is the pathway for people under 15? | WHO does not yet recommend CAD below 15, so an alternative route must exist by design. |
The honest limits
CAD for chest X-ray is a screening and triage aid under a conditional, low-certainty recommendation. It is not a diagnostic test, does not establish TB, and does not remove the need for bacteriological confirmation. Its recommended scope stops at people aged 15 and older. Its performance varies between settings, which is the stated reason WHO asks for local calibration instead of a single global cutoff. Screening yield also depends on the underlying prevalence in the group screened, so a result observed in one district does not transfer unexamined to another.
Where this fits with Clinoble's work
Clinoble Innovations Private Limited (Hyderabad, Telangana, India) builds decision-support and cascade-tracking tools for this class of problem. Two Indian patent applications in quantitative clinical imaging are pending: IND 202641088451 and IND 202641090786. Nothing below is endorsed by WHO, NTEP or any government body, and nothing below replaces confirmatory molecular testing.
TB F.I.R.S.T is a cascade tracking progressive web app aimed at the handoff problem above: recording who was screened, who was flagged, who was referred for confirmatory testing, and whether that referral closed with a documented outcome. It does not interpret images and does not produce a diagnosis.
CardioPulmo works on respiratory and cardiac sound patterns as a separate triage signal alongside imaging-based screening. The governing principle is the same one that applies to CAD: a flag is a reason to evaluate further, not a result to act on alone.
For teams weighing whether a screening tool fits their setting, clinical evaluation services covers study design, evidence review and the documentation trail behind claims.
Clinoble Innovations Private Limited, Hyderabad, Telangana, India. Founder and Director: Dr. Jaideep Rao M., MBBS, MD Community Medicine.