AI-Driven Geometric Modeling and Analysis for Medical Imaging and Scientific Data

Abstract

Understanding and modeling complex geometric structures from medical imaging and scientific data remain central challenges in computational science and biomedical engineering. Accurate geometric representations are essential for visualization, quantitative analysis, simulation, and clinical decision making. Yet existing workflows face persistent limitations. Traditional segmentation methods rely heavily on manual annotations and handcrafted priors, limiting their generalizability across modalities. Deep learning-based approaches have improved performance but still struggle with limited annotated datasets, insufficient uncertainty modeling, and difficulties segmenting anatomical structures with substantial scale variation. Moreover, converting voxel segmentations into smooth, simulation-ready meshes often depends on heuristic post-processing, and current surface analysis tools provide limited flexibility for localized geometric characterization in large scientific datasets. Furthermore, the scarcity of high-quality annotated 3D medical data fundamentally limits the scalability and generalization of learning-based geometric modeling methods. This dissertation addresses these challenges through an AI-driven framework that integrates data-driven learning with geometric modeling and analysis. The contributions span three key stages of the geometric pipeline: segmentation, mesh generation, and surface analysis. To complement these stages and mitigate data scarcity, this dissertation further investigates diffusion-based generative modeling for synthesizing anatomically consistent annotated medical volumes, providing structured data augmentation to enhance downstream segmentation and geometric reconstruction. For segmentation, two complementary methods are introduced. LoGB-Net enhances vessel segmentation by strengthening multiscale feature learning and incorporating Bayesian uncertainty estimation. Sli2Vol+ reduces annotation costs by leveraging a slice-to-volume correspondence mechanism that propagates sparse 2D annotations into coherent 3D segmentations, enabling annotation-efficient volumetric reconstruction. For mesh generation, AortaDiff introduces a diffusion-based generative model conditioned on CT volumes to synthesize anatomically consistent aorta meshes. By jointly modeling centerline geometry and local radius variation, AortaDiff reconstructs smooth, continuous, and CFD-ready vascular surfaces. For surface analysis, SurfPatch presents a patch-based representation that decomposes complex surfaces into localized regions for fine-grained morphological characterization and scalable comparative analysis across heterogeneous datasets. Collectively, these contributions advance geometric modeling across three complementary directions: medical image segmentation, diffusion-based mesh generation, and localized surface analysis. In addition, diffusion-driven data augmentation is integrated into this pipeline to enrich structural diversity and improve the robustness of learning-based geometric modeling. By unifying learning-based segmentation, generative geometric reconstruction, patch-level surface characterization, and structured generative augmentation, this dissertation improves the accuracy, efficiency, and reliability of computational analysis in medical imaging and scientific data.

Ähnliche Arbeiten