Machine Learning for Molecular Imaging and Medicine of the Future – AAIMME

Positron Emission Tomography (PET) is a powerful functional medical imaging modality. A radiotracer is injected to the patient, targeting a biochemical process and then gamma ray pairs are detected. Image reconstruction is needed to visualize and quantify the functional process of interest. The quality of the reconstructed image highly depends on the spatial and time-of-flight resolutions of the gamma detectors. The development of ultrafast detectors and dedicated reconstruction tools are today a major challenge, to improve detectability, reducing radiation dose or the exam duration.
The AAIMME project aims to propose reliable Artificial Intelligence (AI) tools optimizing the ultrafast detection chain and the image reconstruction. This work will contribute to a more precise diagnosis and therapeutic monitoring of patients.
On the detector side, innovative density neural networks will provide accurate spatiotemporal localization of the gamma interaction together with its uncertainties. We will study the robustness of the algorithms and their sensitivity to domain changes.
We will develop hybrid image reconstruction methods combining advanced optimization tools, Bayesian modelling, and deep neural network frameworks. These approaches will benefit from fast execution on modern computing environments, from robustness guarantees to the possibly imperfect detection AI-based outputs, and from quantitative uncertainty assessment, for reliable medical decision-making.
We will write dedicated realistic Monte Carlo simulations of the whole PET system to train and validate our AI-based approaches. We will then assess their performance on real data acquired on an experimental bench representative of PET conditions.
To reach these ambitious goals, the consortium gathers partners with expertise in all the fields concerned, working in synergy: researchers in Machine Learning and algorithms, physicists specialized in medical imaging, detector physicists and experimental engineers.