Exploring the lifecycle of artificial intelligence – deep learning for 1D to 4D biomedical data

This thesis, titled "Exploring the Lifecycle of Artificial Intelligence - Deep Learning for 1D to 4D Biomedical Data" presents a comprehensive exploration of the Artificial Intelligence (AI) lifecycle as applied to various domains within the life sciences and medicine, specifically focusing on the application of Machine Learning (ML) and Deep Learning (DL) techniques to data ranging from one to four dimensional data. The study is structured around the AI lifecycle, which is elaborated upon in Chapter 2. This lifecycle encompasses the critical stages of problem definition, data acquisition, model training, evaluation, interpretation, and deployment, offering a holistic approach to the development and implementation of AI systems in medical contexts. A significant portion of the thesis delves into the interpretation of Electrocardiography (ECG) data, as detailed in Chapter 3. This chapter introduces the PTB-XL database, a large-scale dataset for ECG analysis, and discusses the application of eXplainable Artificial Intelligence techniques to enhance model transparency and trustworthiness. The study demonstrates how XAI can be employed to make AI models more interpretable and clinically relevant, which is crucial for their adoption in medical diagnostics. The thesis then shifts focus to the segmentation of 2D histological images, which is critical for accurate medical diagnostics, particularly in dermatology. As described in Chapter 4, this work explores advanced semantic segmentation techniques to identify and classify different biological structures within a Whole Slide Image (WSI). The chapter highlights the challenges and potential solutions in implementing DL models for this purpose, emphasizing the importance of precise segmentation in improving diagnostic accuracy. Further expanding on the dimensional complexity, Chapter 5 addresses the analysis of 3D neutrophil cell nuclei shapes using confocal fluorescence microscopy. This chapter provides insights into the cellular morphology, employing feature extraction, dimensionality reduction, and clustering methods to understand the variations in neutrophil shapes. The analysis underscores the potential of AI in uncovering morphological features that are not easily discernible through traditional methods. The final application discussed in this thesis involves the analysis of 4D lymphocyte dynamics in human lymphoid tissue, as outlined in Chapter 6. This chapter examines the temporal dynamics of lymphocytes using time-lapse confocal microscopy, demonstrating how AI can be utilized to gain novel insights into cellular behavior over time. The work presented here highlights the significant contributions of AI to understanding complex biological processes, potentially leading to new discoveries in immunology. In conclusion, this thesis not only advances the state-of-the-art in the application of AI to the life sciences but also provides a robust framework for the systematic development and deployment of AI models in medical research. By adopting the AI lifecycle approach, the research aims to address each stage, from problem definition to deployment, in a comprehensive manner, with the goal of developing AI solutions that are not only technically sound but also have potential for clinical applicability. The findings of this thesis suggest potential implications for the future of AI in medicine, particularly in enhancing the interpretability, transparency, and trustworthiness of AI models, which are essential for their integration into clinical practice.

Author dc:creator

• Wagner, Patrick

Advisor dc:contributor.advisor

• Samek, Wojciech

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-sa/4.0/

Language dc:language.iso

en

Identifier URI

https://doi.org/10.14279/depositonce-23016

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/24202