Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

ST: Fachverband Strahlen- und Medizinphysik

ST 4: AI in Medicine

ST 4.1: Talk

Wednesday, March 20, 2024, 09:30–09:45, PC 203

Enhancing Brain Tumor Characterization through Machine Learning on Raw MRI K-Space Data — •Marco Schlimbach¹, Jens Kleesiek², Kevin Kröninger¹, Moritz Rempe², and Jens Weingarten¹ — ¹TU Dortmund University — ²Institute for AI in Medicine (IKIM)

Differentiating brain tumor types is a diagnostic challenge, influencing subsequent therapeutic decisions. Clinics employ various methods to examine tumors, including biopsies and Magnetic Resonance Imaging (MRI). One goal of research is the improvement of MRI-based tumor characterization to avoid invasive biopsies. Numerous Machine Learning (ML) approaches have been developed to enhance tumor classification in MRI scans. However, most of these methods use the final images created by MRI scanners through a complex reconstruction pipeline, involving filtering operations. This process, aimed at producing human-interpretable images, loses phase information of the complex-valued raw MRI data, which potentially has diagnostic value.

In an ongoing data acquisition process from our medical partners, raw data, known as k-space data, is extracted from clinical MRI scans to create a new unique dataset. The k-space is hypothesized to reveal new features for the characterization of tumors. This work introduces the innovative approach of applying ML techniques directly to this MRI k-space data. It aims to utilize the complete information of the raw MRI data by using different complex-valued neural network architectures. A generative neural network approach is introduced with the objective to produce synthetic k-space data.

Keywords: AI; MRI; k-Space; Machine Learning; Tumor Classification