DEEP LEARNING IN NUCLEAR MEDICINE AND MOLECULAR IMAGING FOR IMPROVED CLINICAL OUTCOMES

This thesis aims to develop deep learning (DL) approaches for medical image analysis tasks in nuclear medicine. We introduce novel DL methods for automated classification, segmentation, and prognostic tasks to improve clinical outcomes in oncology and neurology. For oncological applications, we focus on DL-based methods for tumor classification and segmentation on positron emission tomography (PET) imaging. We present an interpretable DL and radiomics framework for automated classification of prostate cancer on prostate-specific membrane antigen (PSMA)-targeted PET. The framework takes convolutional neural network-extracted features, radiomic features, and tissue type information extracted from a PET image as inputs and outputs a predicted PSMA reporting and data system (PSMA-RADS) category. The framework provided reliable lesion-level and patient-level PSMA-RADS classification on PSMA PET images. The framework was interpretable and provided a confidence score that can assist physicians in making more informed clinical decisions. We propose a modular physics-guided DL-based PET tumor segmentation framework that addresses the challenges of limited spatial resolution, high image noise, and lack of annotated clinical training data. The modular framework generates PET images containing highly realistic simulated images with known ground truth using a stochastic kernel density estimation and physics-based approach. A modified U-net is trained on these simulated images to learn the tumor segmentation task and fine-tuned using clinical data. The framework yielded reliable automated tumor segmentation on 18F-fluorodeoxyglucose (18F-FDG) PET images of patients with lung cancer using a small-sized clinical training dataset, generalized across scanners, and demonstrated the ability to segment small tumors. For applications in neurological disorders, we focus on DL approaches for motor outcome prediction of patients with Parkinson’s disease (PD) using dopamine transporter imaging with 123I-ioflupane (DaTscan). We develop a three-stage, deep learning, ensemble approach for prognosis in patients with PD. The approach extracted features from DaTscan imaging and clinical measures of motor symptoms and trained an ensemble of deep neural networks on different subsets of the extracted features to predict patient outcome. The approach outperformed individual networks trained on subsets of the data, showed promise for longitudinal prognostication in PD, and demonstrated the synergy of imaging and non-imaging information for the prediction task