EigenOrbit Develops Quantum-Inspired Method for Robust Medical Image AI

Insider Brief
- EigenOrbit and researchers from the Hospital Maggiore of Trieste and the University of Trieste have developed a quantum-inspired method for adapting medical image classifiers while preserving mathematical guarantees of certified robustness.
- The method uses orthogonal adapters built from reconfigurable beam splitter circuits, reducing trainable parameters while preserving the original network’s Lipschitz bound.
- In tests involving trichoscopic images for alopecia diagnosis, the adapter certified 63.8% of test images at a perturbation radius of 36/255, compared with 31.6% for LoRA and 35.6% for full fine-tuning.
PRESS RELEASE — EigenOrbit, together with dermatologists at the Hospital Maggiore of Trieste and the University of Trieste, has published a method for adapting medical image classifiers to new tasks while keeping a mathematical guarantee that small image changes cannot alter their predictions. The paper, Quantum-inspired orthogonal adapters for certifiably robust medical image classification, is available on arXiv (2611.18775), and the code is open source.
Medical AI models can change their output when an image changes in ways invisible to the eye: a slight rotation, a second photograph of the same area, or recompression. Certified robustness addresses this by computing, for every image, a radius within which no perturbation can change the prediction. The EU AI Act (Article 15) and the U.S. FDA’s draft guidance on AI-enabled devices both address the robustness of such systems.
The difficulty is that certified networks are trained on general image collections, and medical datasets are too small to train them from scratch. Standard fine-tuning methods, including LoRA, change the network’s Lipschitz bound, and the guarantee weakens.
EigenOrbit’s adapters are built from circuits of reconfigurable beam splitter gates, a structure taken from quantum computing. These circuits are orthogonal for any choice of parameters, so the adapted network keeps exactly the Lipschitz bound of the original. A 2048-wide layer needs 11,264 trainable angles instead of about 4.2 million weights. All computations run on classical hardware, and the circuits map directly to shallow quantum circuits.
The team adapted a certified ImageNet network (BRONet) to trichoscopic images from 143 patients for alopecia diagnosis and to images from 57 patients for alopecia areata activity. Results:
- 63.8% of diagnosis test images were classified correctly and certified at radius 36/255, against 31.6% for LoRA and 35.6% for full fine-tuning.
- At a larger radius (108/255), the butterfly adapter still certified 20.4% of images; LoRA and full fine-tuning certified none.
- Classification accuracy of the butterfly adapter (80.4%) stayed close to full fine-tuning (82.5%) and above LoRA (71.5%).
- The adapter trains about 100,000 parameters, over 300 times fewer than full fine-tuning.
- Adversarial attacks (PGD, APGD-T) confirmed the certificates: none succeeded within a certified radius.
To the authors’ knowledge, this is the first study of certified robustness in trichoscopic image classification. Next steps include patient-level validation on larger datasets and certified backbones built entirely from quantum-inspired orthogonal layers.
Links

Keep track of everything going on in the Quantum Technology Market. In one place.
Mohib Ur Rehman
Get the latest research, company news, and market intelligence every week.
MENTIONED IN THE ARTICLE










