Speaker
Description
Most developments in AI for medical imaging target the (semi-)automated analysis of images, these techniques have however also large potential for improving and accelerating the acquisition and for making the scanner lower cost. In monolithic detectors with SiPM array readout positioning can be improved by using a dedicated network for improving the estimation of the transverse position (about 10%) and estimating the DOI inside the detector. Also positioning accuracy at the edges is significantly improved compared to conventional positioning methods. Using these techniques in 16 mm thick standard LYSO results in an intrinsic resolution below 1.3 mm and shall at the system level approximate the 2 mm limit (due to positron accolinearity). Deep learning is also useful from improving TOF estimation in these detectors and especially for sampled waveforms from the SIPMs large improvements (roughly 25%) can be observed in TOF estimation. On the same detector this results in a coincidence time resolution of 140ps (simulation). CNN based methods are also very effective in lowering the dose. These CNN networks are typically trained on low (25 %) and high (100%) dose count images and can upgrade the low count images to equivalent quality as the high dose ones. This enables to reduce the scan time and/or lower the dose administered to patients. The same AI-dose reduction factor can also be used in the future for lowering the number of detectors or reducing the scintillator volume in expensive systems like Total body PET. Here a 50 % cost reduction (25 % of counts) is foreseen without image quality loss.
Deep learning is being implemented at different positions in the image acquisition chain. Interesting is that all these improvements can be easily combined with each other and have the potential to deliver High resolution clinical TOF-PET imaging at lower dose and with a quite reasonable system cost using a smaller amount of standard PET components (LYSO, analog SiPMs).
