Simulated Quantum Kernels Lost All 24 Matched Tests to a Default Classical SVM on Brain MRI and Breast Ultrasound: Journal of Imaging Informatics in Medicine, 21 September 2026
Quentir Medicine Monitor
Evidence-based insights for quantum medicine. Published by Quentir Systems LLC · September 23, 2026.
On 21 September 2026 the Journal of Imaging Informatics in Medicine published a study by four computer scientists in Lahore, Abu Dhabi and Manchester that put two quantum machine-learning classifiers against ordinary statistical tools on brain MRI scans and breast ultrasound images. In all 24 matched comparisons the top-scoring quantum model scored lower than the top-scoring classical model, and simulating the quantum models on ordinary computers took roughly 50 to more than 220 times as long as running the classical one.
The paper, by M. Usman Hashmi, M. Adnan Hashmi, Muazzam Ali and Raheem Sarwar, asks whether quantum kernels improve medical image classification once every model gets the same features, the same training budget and the same protection against data leakage. Its answer, for the settings it tested, is no: a standard RBF-SVM with its default settings outperformed both quantum kernels in every condition. The authors describe their design as a leakage-controlled benchmark, and that design is the reason the study deserves a close reading.
Many earlier papers on quantum classifiers for radiology and pathology reported encouraging accuracy figures. Several compared a tuned quantum model with a weak classical one, or tested on a single split of the data, where a small overlap between training and test images can inflate a score. This study sets out to give both sides the same conditions and then measure what is left.
For a radiology department or a hospital buyer, the result gives a practical reference point. When a vendor or research partner proposes a quantum classifier for imaging, the first question is whether it was compared with a strong, well-configured classical baseline under the same conditions. This paper shows how large the difference can be when that comparison is made carefully. It also lists the controls such a comparison needs: identical inputs, preprocessing fitted on training images only, several random seeds and reported calibration.
What Hashmi, Hashmi, Ali and Sarwar tested on BrainTumorMRI and BreastMNIST
The study used two public image collections. BrainTumorMRI, the main dataset, sorts brain MRI slices into four classes. BreastMNIST, the reference dataset, comes from the MedMNIST v2 collection described in Scientific Data in 2023 and is built from breast ultrasound images sorted into two classes.
Every model received exactly the same input. The authors passed each image through a frozen ResNet18 network pretrained on ImageNet, a common way to turn a picture into a list of numerical features, and then compressed those features to four or eight numbers with principal component analysis. Standardization and compression were fitted on training images only, which kept test-set information out of the fitted preprocessing steps. Each experiment ran with three random seeds, and on BrainTumorMRI the classifiers were trained on budgets of 250, 500 and 1,000 images.
The two quantum models were support vector machines that measure similarity between images through a quantum feature map. One used a ZZ-style map, the design introduced by IBM researchers in Havlíček and colleagues' 2019 Nature paper on quantum-enhanced feature spaces, and the other a Pauli-style map. Both were computed on a noiseless PennyLane statevector simulator, which calculates the exact result a perfect quantum computer would give. The classical side had four entries: logistic regression, a linear support vector machine, an RBF-SVM left at the scikit-learn defaults, and a compact neural network.
How large the gap was: a mean macro-F1 deficit of 0.2230 across 24 seed-level comparisons
The main score was macro-F1, a measure from 0 to 1 that gives each diagnostic class equal weight, so a model cannot score well by getting only the most common class right. Across the 24 matched comparisons, the top-scoring quantum model minus the top-scoring classical model averaged minus 0.2230, and all 24 differences were negative. The paired t-test and the Wilcoxon signed-rank test in the authors' published results were both significant.
In plain terms, the default RBF-SVM averaged a macro-F1 of 0.7215 across the conditions it was run in, while the stronger quantum model, the Pauli-style kernel, averaged 0.486 and the ZZ-style kernel 0.383. The RBF-SVM beat each quantum kernel on its own in all eight dataset, dimension and budget conditions. Even the linear classical models, at about 0.68, stayed far ahead of the quantum kernels.
Quantum pillar: computing. Technology readiness: TRL 3 of 9. The quantum classifiers ran as working software on public benchmark images, which serve as stand-in data, but only in an exact simulation of a quantum computer, with no run on quantum hardware and no test on hospital scan archives or patients.
Why more training images and more features did not close the gap
Two findings run against a common expectation. Quantum feature maps are often promoted because they can represent data in a very large mathematical space, so more input features should help them. Here the reverse happened. Averaged over the eight conditions on both datasets, moving from four to eight compressed features widened the gap from minus 0.2003 to minus 0.2444, and the simulation runtime penalty grew from about 89 times to about 182 times; on BrainTumorMRI alone the gap widened from minus 0.2027 to minus 0.2666.
More training data helped the quantum kernels more than the classical one, yet the gap stayed wide at every budget tested. On BrainTumorMRI with eight features, the Pauli-style kernel rose from 0.4179 to 0.5597 as the budget went from 250 to 1,000 images, while the RBF-SVM moved from 0.7473 to 0.7640. Three budgets cannot show where the two curves would meet, and the study does not claim to.
The study also reports calibration, meaning how well a model's stated confidence matches how often it is right. The quantum kernels had slightly lower expected calibration error, while the RBF-SVM had better Brier scores and log-loss, two measures of overall probability accuracy. On those two measures the classical model's probabilities were the more accurate ones in this benchmark; the study did not test how clinicians would use either.
This result fits a wider pattern. In 2024, Joseph Bowles, Shahnawaz Ahmed and Maria Schuld tested 12 popular quantum machine-learning models on 160 datasets and found that out-of-the-box classical models outperformed the quantum classifiers overall. The new paper brings the same discipline to medical images. This Monitor has seen the problem from the other side in a heart-disease classifier whose 90.26 percent accuracy came from a simulator, where a headline figure said little about how the model would compare with a tuned classical rival.
What the paper leaves open for radiology: real hardware, patient-level leakage and external cohorts
The authors list their limits clearly. The quantum kernels ran in a noiseless simulation, which represents an ideal case; the effect of hardware noise and a finite number of measurements on classification was left untested, and the runtimes describe simulation on classical computers, not the speed of a quantum computer. The RBF-SVM was left at default settings without tuning, so a tuned classical model could widen the gap further. Only four and eight feature dimensions were tested, only fixed feature maps were used, and only one frozen feature extractor.
The leakage audit checked for duplicated raw images and could not rule out leakage at the level of patients or examinations, since the public datasets do not always identify which slices come from the same person. The data are public benchmarks, and the authors say that any clinical claim would first need testing on external clinical cohorts. They leave open whether richer feature maps, noise-aware kernels or real hardware could narrow the gap.
The benchmarking lesson carries beyond medicine. A similar question arises in defense, where this Monitor's sister lane followed a hybrid quantum classifier applied to radar satellite data: every such claim depends on the strength of the classical model it is compared with.
How Quentir Reads It
For hospitals and imaging vendors, this study moves the burden of proof. A proposal for quantum machine learning in radiology should now show a comparison against a default or tuned RBF-SVM, logistic regression and a modern neural network, fitted on identical features, with training-only preprocessing, several random seeds and reported calibration. A single accuracy figure against an unnamed baseline no longer meets the standard this paper sets.
For patients, nothing in today's care changes. Brain tumor and breast imaging continue to rely on radiologists and on classical software, and this paper gives no reason to expect a quantum classifier to join that workflow soon.
For research funders, the paper is useful as a template. Its protocol can be applied to any future quantum imaging claim, including ones run on real hardware. This Monitor will return to the question when a quantum kernel beats a strong classical baseline under the same controls, on data from a hospital that did not supply the training set.
Sources
Primary source: M. Usman Hashmi, M. Adnan Hashmi, Muazzam Ali and Raheem Sarwar, Journal of Imaging Informatics in Medicine, published online 21 September 2026 (open access), with the authors' results tables deposited on Zenodo, from which the per-condition figures here are taken. Also drawn on: the ebiotrade (生物通) summary of the paper of 23 September 2026, Havlíček and colleagues in Nature (2019) on quantum feature maps, Bowles, Ahmed and Schuld on quantum machine-learning benchmarks (arXiv, 2024), and Yang and colleagues on MedMNIST v2 in Scientific Data (2023); the readiness assessment is this Monitor's own.