Doctoral Defense

Detecting Early Signs of Recovery of Consciousness and a Computational Model on Working Memory

Xi Cheng

December 5, 2024
9:30 AM
Light Engineering, Room 250
Advisor: Sima Mofakham

Ultra-high-resolution positron emission tomography (PET) imaging systems require numerous readout channels, increasing scanner complexity, power consumption, and cost. To address this, we propose a 4-to-1 pixel-to-application-specific integrated circuit (ASIC) multiplexing scheme. This scheme, based on the segmented light guide array design of the Prism-PET detector module, exploits its distinct light-sharing pattern to reduce channels while minimizing performance loss. We also introduce data-driven techniques for demultiplexing the encoded ASIC signals, including a convolutional neural network (CNN) and a non-parametric decision tree model. In addition, we developed an updated version of conventional energy weighted method, the weighted composition-raise-to-the-power approach, which can effectively localize the interaction events on 2D flood histograms without the pretraining. Accurate localization of annihilation photons is critical for PET image reconstruction. However, Compton scattering, where photons transfer energy across multiple crystals (inter-crystal scattering, ICS), degrades the line of response (LOR) accuracy, leading to lower image quality. To address this, we implemented a deep neural network with an autoencoder architecture to recover ICS event locations. The model was first trained on simulated data and later adapted to experimental data by generating synthetic training sets that closely resemble real conditions. We further present a novel platform for PET image reconstruction, built on the Differentiable Rendering Just-in-Time (Dr. JIT) compiler. This platform integrates physically-based differentiable rendering (PBDR) into the reconstruction process, providing comprehensive physical modeling. Dr. JIT executes millions of parallelized Monte Carlo integrations within a "megakernel," significantly reducing memory usage and inter-kernel communication. This efficient handling of high-dimensional problems enables the integration of complex models without significantly increasing computational costs, while simultaneously producing higher-resolution images. By combining multiplexing design, ICS event recovery, and inverse rendering (IR)-based techniques, this work enhances PET image reconstruction, improving diagnostic accuracy and potentially leading to more reliable clinical outcomes.