Monte Carlo simulation tools will be adapted to simulate the specific geometry of the ProVision prostate cancer imaging. Geometries and timing resolutions will be integrated into the simulation framework to predict its performance in terms of lesion detectability. Currently, there is no image reconstruction software focused on using wide-angle tracks with a coincidence time resolution of about 100 ps. The development of this software is critical for the exploitation of ProVision organ-specific PET scanner and improving the sensitivity and specificity. The reconstruction from under-sampled PET data will be investigated through various strategies for data recovery involving not only dictionary learning based in-painting techniques, but also ground-breaking techniques based on compressed sensing and deep learning algorithms using generative adversarial networks.
In this regard, novel compressed sensing algorithms adjoined with the reconstruction step both in projection space, as a pre-processing step prior to reconstruction, and in image space will be implemented and evaluated. The aim is to develop a ground-breaking algorithm for constant optimization of the algorithms where a generative adversarial network will learn from clinical feedback extracted from patients’ history and outcome including but not limited to biopsy. In this way, the deep learning network will constantly move state of the art forward as a deep learning network, improving with each new feeding of the training data. An automated test protocol to test-retest the algorithms, which will also be used for software verification prior to implementation and clinical adoption will be developed. ProVision will not be integrated with an anatomical imaging modality (e.g. x-ray CT). Therefore, CT-based attenuation correction will not be possible. To enable quantitative molecular imaging, alternative strategies for attenuation correction will be explored. The aim is to develope a constrained maximum likelihood reconstruction of activity and attenuation as well as hybrid atlas-emission-based algorithms for attenuation correction in pelvic imaging.
For attenuation estimation, UniGe lab propose to use a constrained quadratic soft-tissue preference prior to suppress large deviation of attenuation coefficients of soft-tissue from their expected value. Moreover, a novel deep convolutional neural network model with highly accelerated convergence rate suitable for applications where the number of training subjects is limited to generate accurate synthetic CT images will be used. The proposed accelerated convolutional neural network (ACNN) employs the visual geometry group (VGG16) model without fully connected layers combined with residual network implemented in an encoder-decoder structure. This workpackage will also include thorough simulation and preliminary clinical studies that will be used to assess the performance of the proposed reconstruction and quantitative imaging methodologies. To quantitatively evaluate the performance of the reconstruction methods with respect to a ground truth, realistic Monte Carlo simulations will be performed using compu tational pelvic phantoms and a model of the geometry of the ProVision PET scanner with high temporal resolution.