DEVELOPING NOVEL IMAGING-BASED CLINICAL OUTPUT PROGNOSTICATION MODELS AND IDENTIFICATION OF IMAGING BIOMARKERS IN OVARIAN CANCER PATIENTS
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Selecting an effective treatment regimen is central to cancer management. Although clinicians consider numerous patient- and disease-specific factors, many biologically relevant characteristics remain inaccessible through conventional clinical assessment, leading to suboptimal treatment decisions and variable patient outcomes. Medical imaging offers a noninvasive means of capturing tumor phenotype and heterogeneity, enabling the development of quantitative imaging biomarkers that may improve personalized treatment selection. This dissertation investigates imaging biomarkers derived from computed tomography (CT) and digital pathology, spanning handcrafted radiomics, deep learning, and foundation model representations, for predicting six-month progression-free survival (PFS) following platinum-based chemotherapy in ovarian cancer. The first study investigated whether incorporating volumetric information improves radiomics-based prediction by comparing handcrafted features extracted from a single two-dimensional (2D) tumor slice with those extracted from the full three-dimensional (3D) tumor volume. More than 1,200 radiomics features were extracted and evaluated within a machine learning framework incorporating feature selection and support vector machine classification. While 3D radiomics demonstrated improved predictive performance compared with conventional 2D analysis, the results also indicated that incorporating the entire tumor volume introduced additional feature redundancy and variability. Building upon these findings, the second study proposed an incomplete-3D radiomics strategy designed to balance volumetric information with feature stability. Rather than analyzing the entire tumor volume, radiomics features were extracted from the central tumor slice together with a limited number of adjacent slices. This compact volumetric representation consistently outperformed both conventional 2D and full 3D radiomics. Feature redundancy analysis further demonstrated that the most discriminative imaging biomarkers originated from the central tumor region, suggesting that limited volumetric context captures the majority of clinically relevant information while reducing redundant image characteristics. The third study investigated complementary CT imaging biomarkers derived from both handcrafted radiomics and foundation models. Handcrafted radiomics features were combined with deep representations generated by Vision Transformer (ViT) and MedSAM models from metastatic tumor CT images. A nested cross-validation framework with in-fold feature selection and model optimization was employed to compare individual and fused imaging biomarkers. The results demonstrated that foundation model representations provided highly informative imaging biomarkers, while their integration with handcrafted radiomics produced robust and interpretable predictive models. Model interpretation further revealed that MedSAM-derived features contributed most strongly to treatment response prediction, whereas selected radiomics descriptors captured complementary measures of intratumoral heterogeneity. Recognizing that chemotherapy response is influenced by both the intrinsic biological characteristics of the primary tumor and the phenotypic behavior of metastatic disease, the fourth study developed a multimodal imaging framework integrating digital pathology and radiology. Histopathological representations were extracted from primary tumor whole-slide images using the CONCH foundation model, while complementary imaging representations were obtained from metastatic CT lesions using MedSAM. Multiple multimodal fusion strategies, including early fusion, late fusion, and attention-based fusion, were investigated to learn complementary biomarkers spanning microscopic tissue architecture and macroscopic metastatic phenotypes. By integrating imaging biomarkers from both disease compartments, this biologically motivated framework captures complementary aspects of ovarian cancer biology that cannot be characterized by either modality alone and demonstrates the potential of multimodal foundation models for improving treatment response prediction. Overall, this dissertation demonstrates the evolution of imaging biomarkers from handcrafted radiomics to multimodal foundation model representations for treatment response prediction in ovarian cancer. The findings show that compact volumetric radiomics improves robustness over conventional 3D analysis, foundation model-derived imaging representations provide powerful predictive biomarkers, and integrating pathology-derived information from the primary tumor with radiology-derived information from metastatic lesions offers a promising strategy for capturing complementary aspects of tumor biology associated with chemotherapy response. Together, these studies establish a scalable framework for multimodal imaging biomarker discovery and lay the foundation for future integration with genomic, molecular, and clinical information to advance precision oncology.