Open Collections

Abstract

Modern nuclear medicine imaging pipeline involves image generation, enhancement, and analysis, each facing challenges in reconstruction fidelity, quantitative reliability, and interpretation. This thesis presents deep-learning approaches to overcome these limitations throughout the nuclear medicine pipeline. In Chapter 2, we propose DIP-SPECTNet, an unsupervised approach leveraging the inductive bias of convolutional networks to denoise SPECT projections without paired training data. Exploiting the deep image prior technique, our method separated noiseless photopeak projections from Poisson noise while preserving anatomical features in low-count regimes. In Chapter 3, we present DAWN-FM, a novel framework for solving ill-posed inverse problems through data and noise-aware flow matching. Incorporating embeddings for measured data and noise characteristics into the training process, we solved deblurring and tomography inverse problems in the presence of noise. Our method’s ability to sample from the learned posterior enables the exploration of the solution space and facilitates uncertainty quantification. In Chapter 4, we develop a comprehensive framework for evaluating deep-learning methods for lymphoma quantitation. Rigorous comparison with expert annotations showed that networks match physician performance for large lesions while revealing shared limitations in detecting small, low-contrast abnormalities. In Chapter 5, we introduce IgCONDA-PET to overcome annotation scarcity for training networks for anomaly detection in PET. Our weakly-supervised approach combined attention-mechanisms with counterfactual diffusion modeling to localize lesions without pixel-level supervision, outperforming other competing methods across diverse cancer phenotypes. In Chapter 6, we propose Thyroidiomics, a machine-learning framework for thyroid disease classification using scintigraphy imaging. Integrating deep segmentation with radiomics analysis achieved physician-level accuracy while reducing additional test requirements and assessment time. Finally, in Chapter 7, we present Multiscale Stochastic Gradient Descent, addressing computational challenges in high-resolution network training. The computation of the gradient of loss using a coarse-to-fine strategy with novel mesh-free convolutions enabled efficient convergence while maintaining resolution consistency, which is crucial for training deep-learning models where training compute is often the bottleneck, especially in the case of high-resolution imaging. Together, these AI solutions have the potential to enhance nuclear medicine from acquisition to diagnosis by addressing core challenges in data quality, annotation needs, and computational efficiency, bridging innovation with clinical implementation.