When people think of Lekton.AI, the first thing that comes to mind is generative AI: the document platforms, legal case analysis, 231 compliance. It is the heart of what we do. But artificial intelligence is a much wider territory than text generation alone, and part of our research explores it in directions that have little to do with generation. Two lines, in particular, are less visible but strategically important: classical deep learning applied to medical imaging and Quantum Machine Learning. In this article we tell them together, because they meet on a concrete, high-impact problem: the early diagnosis of Parkinson’s disease.

The clinical problem

Parkinson’s disease is a neurodegenerative disorder linked to the progressive loss of dopaminergic neurons. One of the most pressing clinical challenges is early, objective diagnosis: by the time the motor symptoms become evident, much of those neurons are already compromised. Functional imaging of the dopaminergic system — for example PET with the C-DOPA tracer, which measures striatal uptake — offers a valuable quantitative signal, but interpreting it requires specialist expertise and lends itself to support from models capable of reading complex patterns in three-dimensional images.

This is where deep learning can contribute: not to replace the clinician, but to provide an additional, reproducible and measurable analysis tool.

BrainNet: classical AI at work on the images

BrainNet is a 3D convolutional neural network for the classification of C-DOPA PET images aimed at the early diagnosis of Parkinson’s. It is the re-engineering of an original research notebook according to current best practices in deep learning for medical imaging — and it is precisely this methodological discipline, rather than the architecture itself, that represents the value of the project.

On the technical side, BrainNet adopts a three-dimensional DenseNet121 (with a lighter baseline for comparison), a pipeline built to be leak-free — that is, free from contamination between training and test — and validation with 5-fold cross-validation, stratified and grouped by patient (StratifiedGroupKFold). The metrics do not stop at accuracy: sensitivity, specificity and AUC are reported with 95% confidence intervals via bootstrap, because on a limited cohort a single number does not tell the truth.

And here lies the most important point, which is worth stating frankly. With a sample of about 46 patients, the realistic goal is not perfect accuracy — a result of that kind, in the original code, was in fact an artefact of data leakage — but rather an honest estimate, complete with uncertainty, of the cross-validation performance. For robust, clinically reliable results, more subjects are needed and, ideally, external validation on an independent cohort. Stating this is not a weakness: it is the condition for doing serious research on health data.

Health data: confidentiality first

Clinical data is a special category under Article 9 of the GDPR. For this reason BrainNet keeps a clear separation between code and data: the DICOM images and patient lists do not reside in the repository, work is done on a pseudonymised CSV (identifier and label only), and the DICOMs must be further de-identified by removing the tags that contain personal information. Their handling remains subject to the study’s ethical approval. It is the same principle we apply in the regulated sectors we operate in: technology advances, but respect for the individual’s data is not negotiable.

Beyond the classical: the quantum route

The second line of research asks: what if part of the feature extraction could take place on a quantum circuit? Quantum Machine Learning studies hybrid quantum-classical neural architectures, in which quantum convolutional layers or parametrised circuits sit alongside traditional networks, with the hypothesis of richer representations or greater data efficiency — an aspect that is particularly interesting precisely when cohorts are small, as in the medical domain.

Our repository qh_algorithms tackles the question in the way we consider correct: with a rigorous experimental comparison between a classical CNN, a hybrid quantum-classical CNN (HQCNN) and a purely quantum reference. Not a single lucky experiment, but a campaign of 10 independent seeds per architecture, with Wilson confidence intervals, paired Wilcoxon tests and bootstrap to establish whether the observed differences are statistically meaningful. All of it built on Qiskit and PennyLane and validated on the EuroSAT image dataset, with a dedicated study of noise resilience — because today’s quantum hardware (NISQ) is noisy, and ignoring that means telling ourselves fairy tales.

It must be said honestly: today this is methodological research on the classification of generic images, not a quantum model already applied to Parkinson’s diagnostics. But the bridge is explicit. The same rigorous-comparison infrastructure — architectures set against each other with serious statistics and attention to noise — is precisely what would be needed to assess, tomorrow, whether and when a hybrid quantum approach brings a real advantage over classical deep learning alone on medical images like those of BrainNet.

The common thread: method before hype

Classical or quantum, generative or predictive, what matters to us is always the same: rigorous validation, declared uncertainty, contamination-free pipelines, data protection, and the competent person at the centre of the decision. It is the same non-negotiable principle of our document platforms — every output declares how it was produced and remains subject to validation by the professional — applied here to a different and extremely delicate domain.

Generative AI is today the most visible technology, and rightly so. But artificial intelligence in the service of real problems — health, diagnosis, scientific research — also requires different tools and an engineering culture made of measurement, caution and transparency. It is the direction we are looking in, alongside our daily work.

---

Note. The projects described are research activities. They do not constitute a medical device, nor a diagnostic tool validated for clinical use, nor medical advice. The results reported in the repositories are preliminary and require larger cohorts and independent external validation before any real-world application. For any health matter it is essential to consult a doctor.

The code is open source on our GitHub (BrainNet, qh_algorithms). Want to talk about it with us? Contact us.