Deep learning for positron emission tomography

Positron emission tomography (PET) is commonly used for the detection and diagnosis of cancer. It works by injecting the patient with a positron-emitting radioactive tracer that gathers around cancerous tissue. The emitted positrons annihilate with nearby electrons, producing two back-to-back gamma photons. Scintillation crystals absorb these gamma photons and re-emit them as visible light, which is then detected by silicon photomultipliers (SiPMs). This information can then be used to reconstruct the 3D tracer distribution.

This project focuses on improving the overall performance of PET scanners by substituting or enhancing conventionally used signal processing techniques with modern deep learning algorithms. These algorithms are implemented both at the system level (for image reconstruction and denoising) and the detector level (for positioning and timing of the gamma absorptions). The improved scan quality allows for detection of diseases during earlier stages.

Contact: Jens Maebe and Milan Decuyper