Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · November 11, 2026.
A research team in India has attached a four-qubit quantum circuit to six standard image-recognition networks and tested whether it helps detect pulmonary tuberculosis on chest X-rays. The top-scoring hybrid reached an area under the curve of 0.9686 on 662 public radiographs, slightly above its classical counterpart at 0.9668.
The study appeared in Frontiers in Artificial Intelligence on 11 November 2026. Its fourteen authors, led by first author Sonet Daniel Thomas with Rajesh Raju and Jalaluddin Akbar Kandel Codi as corresponding authors, work mainly at Yenepoya (Deemed to be University) in Mangalore, India, in its integrative omics data science unit and its medical college. The quantum circuit ran on a software simulator on ordinary processors, with no quantum hardware involved.
The most useful result for clinicians is a control experiment the authors built in themselves. They replaced the quantum circuit with a simple classical stand-in built from edge-detection texture features, kept the same feature-output dimensionality and the rest of the model, and retrained everything on the same data splits. That stand-in reached 0.9658, and the authors' paired statistical test found no significant difference in classification errors between it and the quantum hybrid.
In their comparative analysis the authors write that the improvement "should be regarded as incremental rather than definitive evidence of a quantum-specific advantage". For a hospital or a screening program weighing claims about quantum machine learning in radiology, that sentence is the part of the paper to keep.
What the Yenepoya team built and tested on 662 Shenzhen chest X-rays
The images come from the Shenzhen Hospital chest X-ray set, released by Jaeger and colleagues at the U.S. National Library of Medicine and described in Quantitative Imaging in Medicine and Surgery in 2014. It holds 662 frontal radiographs from Shenzhen No. 3 People's Hospital in China: 326 normal and 336 with tuberculosis. The near-even split makes the scores easier to read, though it does not match the low prevalence of a real screening population.
Each hybrid model has two branches. A classical network pre-trained on everyday photographs, one of VGG16, ResNet50, DenseNet121, InceptionV3, EfficientNetB0 or EfficientNetV2-S, reads the full image at 256 by 256 pixels. In parallel, the same X-ray is shrunk to a grid of two by two pixels. Those four numbers become rotation angles on four simulated qubits, which are linked by three entangling gates and then measured. The few values from the quantum branch are joined to the hundreds from the classical network, and a small final layer gives the probability of tuberculosis.
The methods section describes five-fold cross-validation with fifty training rounds per fold on standard processors. The discussion gives a conflicting account: it says the InceptionV3 hybrid was limited to two rounds per fold because of a memory overflow during gradient computation. They report sensitivity, specificity, accuracy and training time for all six classical networks, all six hybrids and all six stand-in versions.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The method exists as a proof of concept tested on a single public image set, with the quantum part run only on a software simulator, many steps before any trial on real quantum hardware or with patients in a clinic.
How the hybrid compared with the classical EfficientNetB0 and its edge-detection stand-in
The headline numbers are close together. The quantum hybrid built on EfficientNetB0 reached an area under the curve of 0.9686, plus or minus 0.0074 across folds. The plain EfficientNetB0 reached 0.9668, and the stand-in version reached 0.9658. Differences in the third decimal place sit well inside the spread between folds.
The trade-offs at the chosen threshold point in different directions. The hybrid found more true cases, with a sensitivity of 93.2 percent against 91.4 percent for the classical network, and the authors note that missed cases delay treatment and allow further spread. It also raised more false alarms: its specificity was 86.2 percent against 89.3 percent, and its overall accuracy was 89.7 percent against 90.3 percent. The stand-in reached the highest accuracy of the three, 90.6 percent, with a specificity of 91.7 percent.
The paired McNemar test, which checks whether two models make different errors on the same images, separated the hybrid from the plain classical network. Against the stand-in it gave a chi-square of 0.071 and a p-value of 0.789. A different set of errors is a weaker claim than better diagnosis, and the authors' results passage limits these tests to differences in error patterns between paired models.
Training time depended on the network. The EfficientNetB0 hybrid needed about 4.7 hours for the full cross-validation, less than the classical version. The VGG16 hybrid needed about 28 hours, more than twice its classical and stand-in versions, because simulating the quantum circuit's gradients grows costly with large networks.
What WHO's 2021 screening guideline expects of chest X-ray software, and where this study sits
Computer reading of chest X-rays for tuberculosis is already part of public health practice. The WHO consolidated guideline on tuberculosis screening of 2021 conditionally recommends that computer-aided detection software may be used in place of human readers to interpret digital chest radiographs for screening and triage in people aged 15 and older. A program that buys such software needs evidence of accuracy in its own population and a threshold chosen for local conditions.
Measured against that practice, the Yenepoya study is an early methods paper. It uses one public set of 662 images from a single hospital, a binary normal-or-tuberculosis label, and no external test set. The authors list the limits themselves: simulation in place of real quantum hardware, information lost when an X-ray is reduced to four pixels, and no separation of tuberculosis from other lung disease.
Several studies we have covered share this design question. Our note on a University of Florida and MD Anderson review of quantum machine learning for MRI, chest X-ray and CT found that nearly all published systems in radiology are hybrid quantum-classical pipelines tested on small, downsampled image sets or on simulators, and that clinical deployment is not yet feasible. The Yenepoya paper fits that description, and its stand-in adds a direct test of how much the quantum branch contributes.
Which tests would move hybrid quantum tuberculosis detection toward a clinic
Three steps would make the result more useful to a screening program. The first is external validation on radiographs from other hospitals and countries, including the harder cases the authors mention in their introduction, such as children and people living with HIV. The second is a comparison against computer-aided detection products that programs already use, at thresholds set for screening. The third is a run on real quantum hardware with the noise that comes with it, which the authors identify as future work.
Until those results exist, a hospital buyer can treat this paper as a careful negative control for its own field. On one dataset, its quantum branch showed no statistically significant difference in paired classification errors from a classical branch with the same feature-output dimensionality, and the paper reports that finding plainly.
Sources
Primary source: Sonet Daniel Thomas, Rajesh Raju, Jalaluddin Akbar Kandel Codi and eleven co-authors, mainly at Yenepoya (Deemed to be University), Mangalore, in Frontiers in Artificial Intelligence, 11 November 2026. Also drawn on: the Shenzhen chest X-ray dataset description by Jaeger and colleagues (2014), the WHO consolidated guideline on tuberculosis screening (2021), and earlier Quentir coverage.