´óÏ㽶ÊÓƵ

Optional

Abstract: Image-guided therapy enables clinicians to monitor treatment progress, verify accurate delivery, and make immediate adjustments. Among various imaging modalities available for this purpose, optical imaging techniques offer distinct advantages for real-time guidance through their non-ionizing nature and high temporal resolution. However, light propagation in biological tissue fundamentally limits these imaging modalities by obscuring spatial information and introducing noise artifacts. These physical constraints make it hard to accurately interpret subsurface structures and compromise the reliability of real-time guidance.

This thesis develops deep learning methods to overcome these limitations in two distinct optical imaging modalities. The first approach aims to improve time-of-flight fluorescence imaging with a spatiotemporal neural network. Current fluorescence-guided surgery relies on steady state fluorescence images, which lack depth information and are confounded by background fluorescence signals or noise. Time-of-flight fluorescence imaging is a transformative modality that captures temporal delays of emitted photons. These time-resolved profiles can encode depth information. However, recovering true fluorophore topology from these temporal measurements represents an ill-posed inverse problem that traditional analytical fitting methods cannot adequately solve. A spatiotemporal deep learning model is developed to learn the mapping between time-resolved signals and fluorophore distributions. The performance of the model is first demonstrated by Monte Carlo simulated datasets validated against experimental measurements with homogeneous optical setup. This model is subsequently adapted to tissue with heterogeneous optical properties by adding extra surface reflectance profiles as input.

The second approach advances Cherenkov imaging for patient positioning verification. Cherenkov images contain patient-specific bio-morphological features such as vasculature that can serve as intrinsic markers for positioning verification. A transfer learning strategy employing SegResNet was developed to automatically segment vessels in Cherenkov images. Evaluation on clinical Cherenkov datasets demonstrates the model's advantages compared to manual segmentation in terms of processing speed, accuracy, and consistency.

The third approach intends to denoise single-frame Cherenkov images. Conventional Cherenkov imaging requires frame averaging to achieve acceptable image quality, sacrificing the temporal resolution needed for dynamic monitoring. A diffusion-based model is developed for single frame denoising, exploiting the correspondence between temporal frame accumulation and the diffusion process to recover clean images while preserving spatial features.

Thesis Committee: Petr Bruza (Chair), Brian Pogue, Arthur Pétusseau, Yucheng Tang (Nvidia/Vanderbilt)