Positron emission tomography (PET) is a medical nuclear imaging technique widely used in oncology and neurobiology. The decay of the radioactive tracer emits positrons, which annihilate into two back-to-back photons of 511 keV. These pairs of photons are detected in coincidence and used to reconstruct the distribution of the tracer activity in the patient's body.
In this thesis, we propose to contribute to the development of the cutting-edge patented technology ClearMind. The first prototype is currently being tested in the laboratory. The proposed detector uses a monolithic lead tungsten crystal in which Cherenkov and scintillation photons are produced. Those photons are converted to electrons by the photo-electric layer and multiplied in a microchannel plate. The induced electrical signals are amplified by gigahertz amplifiers and digitized by the fast acquisition modules SAMPIC. The opposite surface of the crystal will be equipped with a matrix of the silicon photo-multiplier. Machine-learning techniques will be applied for processing the complex acquired signals in order to reconstruct the time and coordinates of the gamma-conversion in the crystal.
The candidate will work on the development of high-efficient machine learning algorithm for the reconstruction of the gamma-conversion vertex in the monolithic crystal. In particular, this work consists in the evolution and improvement of the existing Geant4 detector simulation for its adjustment to the prototype performances as measured in the laboratory. This simulation will provide a training dataset for the development and optimization of deep neural networks with a focus on reconstructing vertex parameters and estimation of the uncertainties on these parameters (i.e., robust IA).
Calibrations on multiple detectors will prepare several batches of realistic performance test data, allowing us to assess the stability of our methods across domain changes. These data inherently contain noise and will thus also serve as rigorous tests of robustness.
These algorithms will enable the efficient reconstruction of gamma interactions using either the full signal shape and/or pre-processed data (feature engineering). Special attention will be given to developing compact, efficient, and fast networks. The possibility of embedding these algorithms in FPGA for real-time reconstruction may also be explored.