Date of Award
12-20-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Deep learning driven computational methods have profoundly impacted medical imaging by advancing feature extraction, and by establishing a critical foundation for downstream image analysis and interpretation. Co-clinical trials are an emerging area of investigation in which a clinical trial is coupled with a corresponding preclinical study to inform the corresponding clinical trial to enhance translational accuracy and therapeutic development. Within this framework, quantitative imaging and pathology plays a fundamental role in both clinical and preclinical arms of co-clinical trials. Quantitative imaging enables enabling identification and validation of imaging biomarkers which are leveraged to predict therapeutic response, characterize disease phenotypes, and inform patient outcomes, while pathology provides complementary insights validating underlying biological process. However, co-clinical imaging and pathology faces several challenges, including limited throughput, suboptimal imaging workflows, quantification uncertainty, lack of reproducible frameworks and constraints in biological interpretability. These challenges highlight the need for developing reproducible and robust deep learning based computational frameworks for co-clinical imaging and pathology. The objective of my thesis is to develop such frameworks and evaluate their performance using multi-objective validation metrics including task-based quantitative metrics to improve interpretability and translation. To address critical challenges in co-clinical imaging and pathology, I have developed three distinct deep learning-based computational frameworks, each optimized for specific tasks within preclinical imaging and clinical pathology. The first framework focuses on development and optimization of automatic pipeline to mitigate the manual efforts, reproducibility issues and inter-observer variability associated with preclinical tumor segmentation in co-clinical studies. The novel DR2U-Net architecture was developed to perform automatic tumor segmentation in multi-contrast MRI, especially optimized for small-animal imaging. The efficacy of the framework was evaluated by reproducibility assessments and quantitative accuracy by incorporating radiomic feature analysis into the pipeline. Furthermore, the framework has been deployed and integrated on the Preclinical Imaging XNAT-enabled Informatics (PIXI) cloud-based platform to enable scalability. The second framework focuses on addressing the challenges in low-count PET (LC-PET) imaging, which could be used to enhance scanner throughput and streamline animal logistics in preclinical PET studies. However, the application of LC-PET imaging is limited by lower SNR, quantification variability and lesion detection difficulty. This tradeoff is addressed by developing a framework to generate Standard-Count PET (SC-PET) from LC-PET utilizing both supervised (ARD-Net) and unsupervised (N2N-MBRNet) learning paradigm. Multi-objective metrics including fidelity-based evaluations, task-based segmentation accuracy, task-based quantification performance, and radiomic feature correlation were applied to validate the performance of the framework. Finally, the last framework was developed for the clinical arm in context of co-clinical pathology to automatically predict HER2 score from H&E images by employing a dual-step deep learning network. First, the novel Corr-A-Net architecture was developed to learn shared correlated feature representations between H&E and IHC. Finally, these H&E generated shared correlated representations are utilized to predict HER2 score using a predictor network. Furthermore, each prediction is accompanied with a confidence score generated by the surrogate confidence score estimation network trained using incentivized mechanism and an interpretable map to validate clinical relevance. Collectively, these frameworks have demonstrated the potential to increase throughput, enhance reproducibility, and streamline imaging workflows, contributing to greater quantitative and biological interpretability within both the preclinical and clinical phases of co-clinical studies.
Language
English (en)
Chair
Kooresh Shoghi
Committee Members
Abhinav Jha; Farrokh Dehdashti; Hongyu An; Richard Laforest