The Short Answer
A battery-powered X-ray that fits in a backpack, paired with AI that reads the chest image, can screen for tuberculosis in a village with no radiologist, and with tools such as Delft Imaging's CAD4TBbox, no internet either. On ibl.ai, health systems apply the same pattern to AI agents: you own all the code and the data and deploy anywhere, including fully offline.
TB made an estimated 10.6 million people ill in 2025 and killed 1.13 million, according to the WHO tuberculosis fact sheet. Many of those people live far from an X-ray room and further still from a radiologist.
ibl.ai does not build X-ray machines or diagnostic software. This post looks at what the portable TB screening programs got right about where AI should run, because the same decision faces every health system deploying AI agents.
What is the backpack X-ray that screens for TB in remote villages?
There are several, and the story circulating online blends them. The one most often named is LeXIS from Lenek Technologies, a Kanpur startup founded in 2021 and incubated at IIT Kanpur's Startup Incubation and Innovation Centre.
LeXIS is a 3.5 kg, battery-powered, ultraportable X-ray system, reports Startuppedia. A licensed radiographer operates it, while AI reads the chest X-ray, generates a report and sends it to a doctor.
Lenek's co-founder Chirag Agrawal demonstrates the device in places a car cannot reach, the same report says. Digital Health News names Lenek's AI radiology platform LIRA and says the initial focus is TB screening where power is unreliable.
Separately, the Indian Council of Medical Research gave IIT Kanpur scientists INR 4.60 crore in 2022 to develop an AI-enabled, battery-operated handheld X-ray, per Digital Health News.
IIT Kanpur's project page says that device "can be packed into backpacks" to reach hard-to-reach populations. Digital Health News reports it is expected to cost less than half as much as imported equivalents.
| Portable TB screening setup | What is published | Where the AI runs |
|---|---|---|
| Lenek LeXIS (India) | 3.5 kg, battery-powered; AI reads the X-ray and sends a report to a doctor | Not published |
| ICMR / IIT Kanpur handheld X-ray (India) | INR 4.60 crore ICMR grant (2022); battery-operated; packs into a backpack | Not published |
| Delft Imaging + CAD4TB (Uganda and 90+ countries) | 17 backpack kits in Uganda since 2020; Delft claims a result in under 5 seconds | On site, fully offline (CAD4TBbox), syncing later |
Does the TB screening AI need an internet connection to read the X-ray?
It depends on the product, and only one vendor in this set says so plainly. Delft Imaging states that its CAD4TBbox analyzes chest X-rays fully offline and syncs results to its cloud once a connection is available.
Delft offers offline, online and hybrid deployment "to match your connectivity", and says the software returns scores, a heatmap and the cardiothoracic ratio in under 5 seconds. It reports CAD4TB deployed in over 90 countries and CE certified as MDR Class IIb.
We found no published statement from Lenek on whether LIRA runs on the device or in the cloud. Its reports are sent to doctors remotely, which needs a connection at some point. A buyer should ask before assuming a village with no signal gets a result on the spot.
Why is AI allowed to replace a human reader for TB screening?
Because WHO said so in 2021. The WHO consolidated guidelines on tuberculosis, module 2: screening, published 22 March 2021, include computer-aided detection software among the recommended screening tools.
The recommendation is conditional, on low-certainty evidence, and covers people aged 15 and older: CAD software may be used in place of human readers to interpret digital chest X-rays for TB screening and triage. It is a screen, not a diagnosis.
Someone flagged by the software still gets a confirmatory test. The WHO fact sheet recommends rapid diagnostic tests, molecular or biomarker-based, as the initial test for anyone with TB signs or symptoms.
What results has backpack X-ray TB screening produced in the field?
Uganda has the clearest public record. Since 2020 the country has invested in 22 digital X-ray systems, in five mobile vans and 17 backpack kits, Gavi's VaccinesWork reported in May 2024.
The field system it describes was developed by Delft Imaging for mobile outreach, and its images are read by CAD4TB. The backpack weighs about 35 kg and travels by car, boat or motorcycle.
Outreach teams screen more than 150 people a day with AI-read X-rays. In 2022 and 2023 the program checked about 50,000 additional people and diagnosed almost 1,000 new TB patients, the same report says.
As of that report, screening and testing reached almost 90% of Uganda's estimated 94,000 annual TB cases, up from 60% five years earlier. The ministry wants an X-ray machine in each of its 135 districts.
What does offline TB screening teach health systems about deploying AI?
Put the model where the work happens, and make the network optional. The portable TB programs work because inference happens beside the patient, while syncing and reporting wait for whatever connection exists.
Most hospital AI is designed the other way round: every request goes to a vendor's cloud, so the AI stops when the link does, and patient data leaves the building on every call. That is a choice, not a law of nature.
The same three questions apply to an AI agent that summarizes a chart, answers a protocol question or trains community health workers. Where does the model run? What happens offline? Who holds the data and the code afterwards?
We covered the hospital version of this argument in why an air-gapped clinical AI platform survives a HIPAA boundary review. The village X-ray is the same principle at the far end of the network.
How does ibl.ai run healthcare AI agents where connectivity is unreliable?
ibl.ai is an agentic AI platform a health system deploys on its own infrastructure: a hospital data center, its own cloud account, or a fully air-gapped network. With the Codebase Transfer engagement, you own all the code and the data, and the platform stays model-agnostic.
That means an open-weight model can run on a local GPU in a district hospital, with no call leaving the site, and the same agents can use a frontier model where policy and bandwidth allow. Switching models does not mean rebuilding the agents.
ibl.ai does not read X-rays and is not a medical device. It is the layer for the AI agents around clinical work, built to sit inside the health system's own boundary. See how it applies to medical and healthcare organizations.