Preview

PDF, English - main document Download (39MB) | Terms of use

Abstract

Owing to high stakes in the field of healthcare, medical machine learning (ML) applications have to adhere to strict safety standards. In particular, their performance needs to be robust toward volatile clinical inputs. The aim of the work presented in this thesis was to develop a framework for uncertainty handling in medical ML applications as a way to increase their robustness and trustworthiness. In particular, it addresses three root causes for lack of robustness that can be deemed central to the successful clinical translation of ML methods:

First, many tasks in medical imaging can be phrased in the language of inverse problems. Most common ML methods aimed at solving such inverse problems implicitly assume that they are well-posed, especially that the problem has a unique solution. However, the solution might be ambiguous. In this thesis, we introduce a data-driven method for analyzing the well-posedness of inverse problems. In addition, we propose a framework to validate the suggested method in a problem-aware manner.

Second, simulation is an important tool for the development of medical ML systems due to small in vivo data sets and/or a lack of annotated references (e. g. spatially resolved blood oxygenation (sO 2 )). However, simulation introduces a new uncertainty to the ML pipeline as ML performance guarantees generally rely on the testing data being sufficiently similar to the training data. This thesis addresses the uncertainty by quantifying the domain gap between training and testing data via an out-of-distribution (OoD) detection approach.

Third, we introduce a new paradigm for medical ML based on personalized models. In a data-scarce regime with high inter-patient variability, classical ML models cannot be assumed to generalize well to new patients. To overcome this problem, we propose to train ML models on a per-patient basis. This approach circumvents the inter-patient variability, but it requires training without a supervision signal. We address this issue via OoD detection, where the current status quo is encoded as in-distribution (ID) using a personalized ML model. Changes to the status quo are then detected as OoD.

While these three facets might seem distinct, the suggested framework provides a unified view of them. The enabling technology is the so-called invertible neural network (INN), which can be used as a flexible and expressive (conditional) density estimator. In this way, they can encode solutions to inverse problems as a probability distribution as well as tackle OoD detection tasks via density-based scores, like the widely applicable information criterion (WAIC).

The present work validates our framework on the example of biophotonic imaging. Biophotonic imaging promises the estimation of tissue parameters such as sO 2 in a non-invasive way by evaluating the “fingerprint” of the tissue in the light spectrum. We apply our framework to analyze the well-posedness of the tissue parameter estimation problem at varying spectral and spatial resolutions. We find that with sufficient spectral and/or spatial context, the sO 2 estimation problem is well-posed. Furthermore, we examine the realism of simulated biophotonic data using the proposed OoD approach to gauge the generalization capabilities of our ML models to in vivo data. Our analysis shows a considerable remaining domain gap between the in silico and in vivo spectra. Lastly, we validate the personalized ML approach on the example of non-invasive ischemia monitoring in minimally invasive kidney surgery, for which we developed the first-in-human laparoscopic multispectral imaging system. In our study, we find a strong OoD signal between perfused and ischemic kidney spectra. Furthermore, the proposed approach is video-rate capable.

In conclusion, we successfully developed a framework for uncertainty handling in medical ML and validated it using a diverse set of medical ML tasks, highlighting the flexibility and potential impact of our approach. The framework opens the door to robust solutions to applications like (recording) device design, quality control for simulation pipelines, and personalized video-rate tissue parameter monitoring. In this way, this thesis facilitates the development of the next generation of trustworthy ML systems in medicine.