Preview

PDF, English - main document Download (8MB) | Lizenz: Relaxing Supervision Requirements for Tomographic Data Analysis with Machine Learning by Zharov, Yaroslav underlies the terms of Creative Commons Attribution - ShareAlike 4.0

Abstract

In this doctoral thesis, the power and potential of advanced imaging techniques, specifically Tomographic Imaging (hereinafter tomography), are explored in an era characterized by the rapid growth of data and the critical need for effective analysis strategies. This work engages with different modalities, such as but not limited to parallel beam X-ray Computed Tomography (CT), and Magnetic Resonance Imaging (MRI). The research is centered around the incorporation of machine learning models, deep learning in particular, to optimize the analysis of tomography scans across various domains, including biology, medicine, and material sciences. This is achieved by navigating the primary challenges associated with the utilization of tomography, namely image preprocessing, data labeling, and model training. This work is organized as a series of chapters, consequently covering those topics in the order in which the proposed techniques would be applied in a practical pipeline of the data analysis.

In Chapter 3 this work explores the applicability of the Noise2Noise denoising technique to the multi-channel imaging datasets, particularly those with significantly reduced Signal-to-Noise Ratio (SNR). Utilizing the self-supervised denoising approach for datasets for biological and material sciences, significant improvements in image quality have been achieved, or, equivalently, the possibility to reduce exposure time has been shown while maintaining image quality.

Chapter 4 of the thesis details the optimization of dataset preparation procedures for training neural networks, specifically concerning tomography segmentation tasks. The study conducted on several openly available medical datasets unravels the critical elements of a useful dataset: quality, diversity, and completeness. It further proposes an optimized labeling procedure that balances these virtues, aiming to deliver the best dataset with minimal effort.

Chapter 5 introduces a novel self-supervised pre-training technique for biomedical tomography called SortingLoss. Its underlying principle is the utilization of the inherent order of slices in a tomography scan volume to pre-train a neural network. This method has been evaluated on medical tomography of lungs affected by COVID-19 and high-resolution full-body tomography of model organisms (Medaka fish) showing lower computational complexity while maintaining results on par with more complex but general approaches.

Lastly, Chapter 6 presents a self-training framework for multi-label segmentation, therefore marking the last stage of the data analysis. The pseudo-labeling method, complemented by a novel Quality Classifier technique to select the best pseudo-labels, and pixel-wise knowledge distillation, has led to improved segmentation performance when tested on the dataset of the Medaka fish brain segmentation.

In sum, this thesis explores the profound potential of integrating advanced computer vision and machine learning tools in the application of tomography imaging. It proposes novel solutions to existing challenges and applies existing techniques in novel circumstances, aiming to remove the burden of manual analysis from the area experts.