Machine Learning and Technologies Development for XEMIS2, the First Liquid Xenon Compton Camera Designed to Push Diagnostic Limits with Functional in vivo Gamma Medical Imaging

The XEMIS (XEnon Medical Imaging System) project, launched in 2004 at the Subatech laboratory, is part of the development of a breakthrough technology in medical imaging. Based on an innovative camera incorporating a monolithic Compton telescope with liquid xenon, this device paves the way for a new generation of imaging systems capable of reducing the radioactive activity administered to patients by a factor of up to 100, without increasing exposure time, marking a major evolution in nuclear medicine practices.

The first XEMIS camera, dedicated to small animal imaging, is currently installed at Nantes University Hospital, with initial experimental data expected in November 2025.

The proposed thesis is part of a strategic collaboration between Subatech (IMT Atlantique), the Data Science department at IMT Atlantique and with the support of CERI SN at IMT Nord Europe. This alliance brings together, within IMT, a rare and already fruitful complementarity: on the one hand, Subatech's expertise in detector physics, in-depth understanding of light/charge chains and flux handling; on the other hand, the skills of the Data Science department and IMT Nord Europe (H. Benhabiles) in deep learning, physics-informed reconstruction and rigorous validation. This collaboration has already borne fruit through recent work in AI applied to detector data and imaging, demonstrating the teams' ability to rapidly produce visible, high-value-added results. Now is the time: the imminent arrival of XEMIS2 data offers a unique window of opportunity for a thesis that closely combines the physics of Compton interactions and the latest generation of AI methods.

With the arrival of the first data, developments specific to this technology will be carried out from start to finish to reconstruct the best possible images. These will cover the fine calibration of sensors and the joint modelling of light and charge responses, the integration of physical constraints (Compton geometry and kinematics, noise regimes) within deep learning architectures, and the implementation of event classification methods based on light information, following on from the work of Giovagnoli et al. 2020 [1, 2] and 2025 [3]. Robust charge/light matching approaches will be pursued in order to increase flow stability, continuing the work initiated in 2025 (M2 internship by L. Le Floch, supervised by the team). The aim is to exploit parsimonious convolutional networks adapted to atypical data with a heterogeneous and sparse spatial distribution, while adopting differentiable programming that explicitly injects Compton physical operators into the core of cost functions and reconstruction pipelines.

Validation will be based on a structured round trip between Monte Carlo simulations and real XEMIS2 data, with metrics relevant to preclinical imaging and a sustained focus on generalisation and uncertainty quantification. The objective is twofold: to demonstrate a tangible gain in image quality at very low doses and to establish a methodology that can be transferred to other gamma/Compton detectors. This integrated approach, driven by the synergy between Subatech, Data Science and IMT Nord Europe, positions IMT as a centre of excellence in physics-informed AI for detection and imaging.

International mobility:
The thesis will include a one-month international mobility period with Prof. E. Barberio's team (Department of Physics, University of Melbourne), in the context of developing a new generation of more sensitive photodetectors. Tests will be carried out on the XEMIS1 prototype (currently at Subatech, transfer to Melbourne in progress), completing the loop of sensors → AI → experimental validation and consolidating the international scope of the project.