Preview

PDF, English - main document Download (17MB) | Lizenz: Translational functional imaging in surgery enabled by deep learning by Ayala Menjivar, Leonardo Antonio underlies the terms of Creative Commons Attribution 4.0

Abstract

Many clinical applications currently rely on several imaging modalities such as Positron Emission Tomography (PET), Magnetic Resonance Imaging (MRI), Computed Tomography (CT), etc. All such modalities provide valuable patient data to the clinical staff to aid clinical decision-making and patient care. Despite the undeniable success of such modalities, most of them are limited to preoperative scans and focus on morphology analysis, e.g. tumor segmentation, radiation treatment planning, anomaly detection, etc. Even though the assessment of different functional properties such as perfusion is crucial in many surgical procedures, it remains highly challenging via simple visual inspection. Functional imaging techniques such as Spectral Imaging (SI) link the unique optical properties of different tissue types with metabolism changes, blood flow, chemical composition, etc. As such, SI is capable of providing much richer information that can improve patient treatment and care. In particular, perfusion assessment with functional imaging has become more relevant due to its involvement in the treatment and development of several diseases such as cardiovascular diseases. Current clinical practice relies on Indocyanine Green (ICG) injection to assess perfusion. Unfortunately, this method can only be used once per surgery and has been shown to trigger deadly complications in some patients (e.g. anaphylactic shock).

This thesis addressed common roadblocks in the path to translating optical functional imaging modalities to clinical practice. The main challenges that were tackled are related to a) the slow recording and processing speed that SI devices suffer from, b) the errors introduced in functional parameter estimations under changing illumination conditions, c) the lack of medical data, and d) the high tissue inter-patient heterogeneity that is commonly overlooked. This framework follows a natural path to translation that starts with hardware optimization. To overcome the limitation that the lack of labeled clinical data and current slow SI devices impose, a domain- and task-specific band selection component was introduced. The implementation of such component resulted in a reduction of the amount of data needed to monitor perfusion. Moreover, this method leverages large amounts of synthetic data, which paired with unlabeled in vivo data is capable of generating highly accurate simulations of a wide range of domains. This approach was validated in vivo in a head and neck rat model, and showed higher oxygenation contrast between normal and cancerous tissue, in comparison to a baseline using all available bands. The need for translation to open surgical procedures was met by the implementation of an automatic light source estimation component. This method extracts specular reflections from low exposure spectral images, and processes them to obtain an estimate of the light source spectrum that generated such reflections. The benefits of light source estimation were demonstrated in silico, in ex vivo pig liver, and in vivo human lips, where the oxygenation estimation error was reduced when utilizing the correct light source estimated with this method. These experiments also showed that the performance of the approach proposed in this thesis surpass the performance of other baseline approaches.

Video-rate functional property estimation was achieved by two main components: a regression and an Out-of-Distribution (OoD) component. At the core of both components is a compact SI camera that is paired with state-of-the-art deep learning models to achieve real time functional estimations. The first of such components features a deep learning model based on a Convolutional Neural Network (CNN) architecture that was trained on highly accurate physics-based simulations of light-tissue interactions. By doing this, the challenge of lack of in vivo labeled data was overcome. This approach was validated in the task of perfusion monitoring in pig brain and in a clinical study involving human skin. It was shown that this approach is capable of monitoring subtle perfusion changes in human skin in an arm clamping experiment. Even more, this approach was capable of monitoring Spreading Depolarizations (SDs) (deoxygenation waves) in the surface of a pig brain. Even though this method is well suited for perfusion monitoring in domains that are well represented with the physics-based simulations on which it was trained, its performance cannot be guaranteed for outlier domains. To handle outlier domains, the task of ischemia monitoring was rephrased as an OoD detection task. This new functional estimation component comprises an ensemble of Invertible Neural Networks (INNs) that only requires perfused tissue data from individual patients to detect ischemic tissue as outliers. The first ever clinical study involving a video-rate capable SI camera in laparoscopic partial nephrectomy was designed to validate this approach. Such study revealed particularly high inter-patient tissue heterogeneity under the presence of pathologies (cancer). Moreover, it demonstrated that this personalized approach is now capable of monitoring ischemia at video-rate with SI during laparoscopic surgery.

In conclusion, this thesis addressed challenges related to slow image recording and processing during surgery. It also proposed a method for light source estimation to facilitate translation to open surgical procedures. Moreover, the methodology proposed in this thesis was validated in a wide range of domains: in silico, rat head and neck, pig liver and brain, and human skin and kidney. In particular, the first clinical trial with spectral imaging in minimally invasive surgery demonstrated that video-rate ischemia monitoring is now possible with deep learning.