Convolutional neural networks for head and neck tumor segmentation in MRI: protocol and sequence optimization

Abstract: Radiation therapy of head and neck tumors requires precise tumor segmentations for an optimal treatment outcome. For treatment planning, tumors are best segmented on magnetic resonance imaging (MRI) data.

Today, convolutional neural networks (CNNs) can automate the tedious segmentation process for some tumors, with an excellent performance:
However, this has not been realized in head and neck cancer so far.
This thesis focuses on the connection of automatic tumor segmentation using CNNs to the MR image acquisition and protocol optimization.
Therefore, the patient setup for head and neck cancer patients is improved to reduce artifacts in diffusion weighted MRI, and the effect of additional distortion correction in ADC parameter maps on the CNN performance is evaluated.
For the segmentation, a feed-forward CNN with 2 pathways for high- and low-resolution features is realized for multi-parametric MRI data of head and neck cancer patients, achieving a segmentation Dice coefficient of up to 65%.
The relative influence of each of 7 MRI input channels is quantified:
It is shown that each of the 7 channels improves segmentation performance of the CNN, and that T2* has the largest overall influence, with a difference in segmentation performance to the reference network of 6%.

Furthermore, a new sequence is developed to simultaneously measure T2 and ADC parameter maps.
The multiecho spin echo sequence with interleaved diffusion blocks is designed to eliminate geometric distortion artifacts present in conventional diffusion weighted MRI using a highly undersampled radial readout in combination with a regularized iterative reconstruction.
The sequence is extensively tested in simulations as well as in MRI phantoms, and first in vivo results are presented

Angaben zum Objekt

• Standort: Deutsche Nationalbibliothek Frankfurt am Main
• Umfang: Online-Ressource
• Sprache: Englisch
• Anmerkungen: Universität Freiburg, Dissertation, 2022

Klassifikation und Themen

• Schlagwort: Head ; Neck ; Kernspintomografie ; Künstliche Intelligenz ; Convolutional Neural Network ; Diffusionsgewichtete Magnetresonanztomografie ; Krebs ; Hals-Nasen-Ohren-Tumor

Beteiligte, Orts- und Zeitangaben

• Ereignis: Veröffentlichung
• (wo): Freiburg
• (wer): Universität
• (wann): 2022
• Urheber: Bielak, Lars
• Beteiligte Personen und Organisationen: Hennig, Jürgen ; Bock, Michael ; Hennig, Jürgen ; Reiter, Günter ; Albert-Ludwigs-Universität Freiburg. Fakultät für Mathematik und Physik

Weitere Informationen

• DOI: 10.6094/UNIFR/226301
• URN: urn:nbn:de:bsz:25-freidok-2263017
• Rechteinformation: Open Access; Der Zugriff auf das Objekt ist unbeschränkt möglich.
• Letzte Aktualisierung: 25.03.2025, 13:49 MEZ