iRegNet: Non-Rigid Registration of MRI to Interventional US for Brain-Shift Compensation Using Convolutional Neural Networks

• Published in: IEEE access 2021, Vol.9, p.147579-147590
• Main Authors: Zeineldin, Ramy A., Karar, Mohamed E., Elshaer, Ziad, Schmidhammer, Markus, Coburger, Jan, Wirtz, Christian R., Burgert, Oliver, Mathis-Ullrich, Franziska
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Brain; Brain-Shift; computer-aided diagnosis; Coordinates; Deep learning; Formability; Image registration; Imaging; intra-operative ultrasound; Magnetic resonance imaging; medical image registration; Medical imaging; Neural networks; Neurosurgery; Registration; Strain; Three-dimensional displays
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2021.3120306

Accurate and safe neurosurgical intervention can be affected by intra-operative tissue deformation, known as brain-shift. In this study, we propose an automatic, fast, and accurate deformable method, called iRegNet, for registering pre-operative magnetic resonance images to intra-operative ultrasound volumes to compensate for brain-shift. iRegNet is a robust end-to-end deep learning approach for the non-linear registration of MRI-iUS images in the context of image-guided neurosurgery. Pre-operative MRI (as moving image) and iUS (as fixed image) are first appended to our convolutional neural network, after which a non-rigid transformation field is estimated. The MRI image is then transformed using the output displacement field to the iUS coordinate system. Extensive experiments have been conducted on two multi-location databases, which are the BITE and the RESECT. Quantitatively, iRegNet reduced the mean landmark errors from pre-registration value of (4.18 ± 1.84 and 5.35 ± 4.19 mm) to the lowest value of (1.47 ± 0.61 and 0.84 ± 0.16 mm) for the BITE and RESECT datasets, respectively. Additional qualitative validation of this study was conducted by two expert neurosurgeons through overlaying MRI-iUS pairs before and after the deformable registration. Experimental findings show that our proposed iRegNet is fast and achieves state-of-the-art accuracies outperforming state-of-the-art approaches. Furthermore, the proposed iRegNet can deliver competitive results, even in the case of non-trained images as proof of its generality and can therefore be valuable in intra-operative neurosurgical guidance.