Graduate Students Seminar
Session Chair: Muhammad Jalil Ahmad
Discussant: Dr. Gobbert
Speaker 1: Ehsan Shakeri
Title Geant4-Simulated 90Y Compton Camera Imaging: 3D Image Enhancement via 3D Attention-Based Residual U-Net
Abstract
Imaging the distribution of Yttrium-90 (90Y) after liver cancer treatment supports estimation of the radiation dose delivered to tumors and surrounding tissues. Compton cameras offer a promising imaging approach, but reconstructed images often suffer from spatial blurring and background artifacts. This talk presents a 3D attention-based residual U-Net for enhancing activity distributions reconstructed from simulated Compton camera measurements. Geant4 Monte Carlo simulations incorporating patient CT anatomy and tumor geometry provide paired reconstructed and ground-truth volumes for training. The neural network combines residual connections, spatial attention gates, and a foreground-weighted loss to emphasize tumor-region accuracy. We introduce the architecture and training strategy, present results demonstrating reduced background artifacts and improved spatial confinement, and discuss remaining challenges in recovering tumor boundaries and quantitative activity.
Speaker 2: Mouhamed Oloude
Title Equivalence Testing for Pooling Correlation Matrices Using a Determinant-Based Criterion
Abstract
In meta-analysis, homogeneity tests are usually conducted to assess whether specific-study effect sizes are adequately represented by a common effect. When homogeneity is rejected, analysts assume that study-specific effects are normally distributed and perform random-effects analysis. However, random-effects analysis is conditioned on normality assumption, or at least large sample sizes where normal approximation can be applied. Furthermore, rejecting homogeneity does not indicate whether individual effect-size - or in this setting, study-specific correlation matrices are sufficiently similar for practical purposes. Therefore, rather than testing equality of correlation matrices, we suggest a test to assess whether correlation matrices are practically equivalent under prespecified practical margins.