Deep Learning Across the Medical Imaging Pipeline: From Disease Phenotyping to Treatment Optimization
Timely initiation of cancer or disease treatment is strongly influenced by the efficiency and reliability of decision-making across the clinical imaging pipeline, from diagnosis and prognosis to therapy planning. Despite recent advances in Deep Learning (DL), important methodological challenges still limit the clinical translation of data-driven approaches. These include the difficulty of learning clinically meaningful representations under imbalanced supervision and hidden stratification during diagnosis, the need to integrate heterogeneous and unpaired multimodal data while preserving biologically grounded contextual interactions for cancer/structure contouring and prognosis, and the rigidity and computational burden of optimization procedures in radiotherapy treatment planning. ; ; This doctoral research develops methodological advances in DL to address these limitations and strengthen algorithmic support along the diagnostic-therapeutic pathway. First, the work proposes representation learning strategies that uncover fine-grained phenotypic variability beyond coarse clinical labels, enabling more robust disease characterization in low-data and hidden-stratification scenarios. Second, it introduces context-aware multimodal learning frameworks designed to operate under multiscale settings and unpaired or missing modalities, supporting reliable segmentation and prognostic modeling by preserving clinically meaningful cross-modal interactions. Third, the thesis explores learning-to-optimize approaches for treatment planning, combining generative modeling and meta-learned optimization strategies to improve the efficiency, stability, and adaptability of traditional planning pipelines under real-world clinical constraints. The proposed methods are developed and validated on public and clinical retrospective datasets spanning computational pathology, radiology, and radiotherapy. Overall, this research aims to contribute methodological insights toward improving scalability, interpretability, and time-to-treatment efficiency in medical imaging pipelines.
Back to Current Students