Deep-learning-based deformable image registration of head CT and MRI scans

• Published in: Frontiers in physics Vol. 11
• Main Authors: Ratke, Alexander, Darsht, Elena, Heinzelmann, Feline, Kröninger, Kevin, Timmermann, Beate, Bäumer, Christian
• Format: Journal Article
• Language: English
• Published: Frontiers Media S.A 05.12.2023
• Subjects: deep learning; deformable transformation; image registration; multimodal; unsupervised
• ISSN: 2296-424X; 2296-424X
• DOI: 10.3389/fphy.2023.1292437

This work is motivated by the lack of publications on the direct application of multimodal image registration with deep-learning techniques for the enhancement of treatment planning in particle therapy. An unsupervised workflow, which seeks to improve image alignment, was developed and evaluated for computed tomography and magnetic resonance imaging scans of the head. The scans of 39 paediatric patients with brain tumours were available. The focus of the two-step workflow, including preprocessing of the scans for normalisation, is deformable image registration (DIR) with a deep neural network, which generates deformation vector fields (DVFs). To obtain a suitable configuration of the network, parameter tuning is performed by varying its parameters, e.g., layer size, regularisation ( λ ) of the DVF and learning rate ( α ). Image similarity was determined with the Dice similarity coefficient, m DSC , using segmented images and the mutual-information metric, m MI . The performance of the deep-learning models was assessed with the inverse consistency, m IC , and the Jacobian determinant, m JD . Inverse consistency is obtained for m IC = 0 mm, while the determinant of a deformed image is expected to be unity. The deep-learning models passed both performance checks, indicated by the mean values m ̄ IC = ( 0.57 ± 1.00 ) m m and m ̄ JD = ( 1.00 ± 0.07 ) . Models with λ ≥ 1 yielded higher m DSC values than models with lower λ values. A small-architecture model with α = 10 –4 was found to be most suitable for DIR, as improvement in image similarity of up to 12% was obtained in terms of m MI . The direct application of deep-learning models produced registered images improving image alignment between scans of different modalities.