United-Imaging-Logo

Quantum AI in nuclear medicine: Can compact models overcome data scarcity?

Quantum artificial intelligence may help nuclear medicine overcome the challenge of limited training data by using compact quantum models to represent high-dimensional medical information more efficiently than conventional deep-learning approaches, though current evidence shows quantum methods perform comparably to traditional algorithms rather than consistently outperforming them.

  • Quantum data encoding converts clinical measurements into quantum states, significantly influencing how models identify patterns in medical images.
  • Nuclear medicine datasets are typically small and heterogeneous, making conventional deep-learning models prone to overfitting due to millions of parameters.
  • Recent studies show quantum and conventional models achieve similar accuracy levels (approximately 70% for quantum versus 69% for conventional methods on PET datasets).
  • Physical quantum computers remain costly and sensitive to noise, limiting practical clinical applications despite theoretical advantages.
  • Clinical validation requires rigorous comparisons with optimized conventional models and external validation before quantum AI can demonstrate meaningful patient benefits.

Vienna -- Quantum artificial intelligence could offer new ways to analyse medical information. But can it help tackle one of nuclear medicine's persistent challenges: limited training data?

At EANM 2026 in Vienna, Dr. László Papp of the Medical University of Vienna presented research examining how clinical data are encoded into quantum states and how this affects cancer classification. His findings suggest that the choice of encoding method can substantially influence a quantum model's performance.

From clinical data to quantum states

First, let's get the technical part straight: Conventional computers process bits representing zeros and ones. Quantum computers use qubits, which can exist in superpositions of states and exhibit entanglement.

Before medical measurements can be processed by a quantum algorithm, they must be mapped into a quantum representation. This process, known as quantum data encoding, can influence how a model processes information and identifies patterns.

While quantum computing is often associated with speed, Papp is investigating another potential advantage: representing high-dimensional medical data using relatively compact quantum systems. The potential of such approaches in nuclear medicine is also discussed in an EANM perspective paper published in 2026.

Why nuclear medicine?

Nuclear medicine datasets are often small, heterogeneous and difficult to combine across specialized centers. Differences in scanners, radiotracers and acquisition protocols can further complicate AI development.

Conventional deep-learning models may contain millions of parameters, increasing the risk of overfitting when only small patient cohorts are available.

Papp is examining whether compact quantum approaches could provide alternatives for selected classification tasks. He distinguishes image segmentation, where many patches can be extracted from each scan, from patient-level predictions, which require sufficiently varied independent patient data.

CLIP 38:39 Papp discusses the statistical challenges of training neural networks with millions of parameters on datasets containing only around 100 patients.

Smaller does not automatically mean better

Quantum states can occupy high-dimensional mathematical spaces, but this does not guarantee better predictive accuracy or easier training. Fewer parameters also cannot create clinical information absent from the source data.

Quantum neural networks can encounter optimization difficulties, including barren plateaus, which can make gradient-based training inefficient. The potential benefit depends on the characteristics of the data and on the model and training procedure, not on model size alone.

What does the evidence show?

Earlier research involving Papp investigated PET radiomics classification with quantum algorithms on simulators and physical quantum hardware.

The 2023 study by Moradi et al. compared methods across PET cancer datasets involving glioma, prostate cancer and lung cancer. In one set of simulation experiments using 16 imaging features, average balanced accuracy was approximately 70% for quantum models and 69% for conventional models. With eight features, conventional methods performed one percentage point better on average.

Selected quantum approaches were also tested on a 21-qubit IonQ Aria computer using error mitigation techniques.

These are experimental, dataset-dependent results rather than evidence of a general clinical advantage.

A 2025 systematic review by Gupta et al., published in npj Digital Medicine, similarly found no consistent evidence that quantum machine learning outperforms conventional approaches across healthcare applications. Of 169 eligible studies, only 16 examined realistic quantum operating conditions involving physical hardware or noisy simulations.

Hardware remains a challenge

Physical quantum computers remain costly and sensitive to noise. Classical simulation helps researchers develop quantum algorithms, but successful simulation does not by itself demonstrate superior speed, cost or clinical accuracy on real hardware.

The 2023 PET radiomics study showed that error mitigation could improve performance on existing quantum hardware, although this does not eliminate the need for rigorous comparisons with conventional algorithms.

Papp's group continues to investigate encoding strategies and small quantum neural networks while tracking hardware progress.

The route toward clinical applications

Potential applications include cancer classification, prognosis and decision support. Demonstrating clinical value, however, will require comparisons with well-optimized conventional models, external validation and evidence of meaningful benefits for patients.

For now, quantum AI in nuclear medicine remains exploratory. Compact representations may offer new ways to analyse imaging data, but whether they can overcome small-data limitations and improve clinical predictions remains unproven.

The question is where, if anywhere, quantum models deliver reproducible benefits beyond established methods.

Page 1 of 133
Next Page