Biomechanical regularization in a deep learning network for a fetal MRI registration and segmentation pipeline

Abstract

Deformable medical image registration and automatic segmentation are essential procedures in many image analysis applications, such as in early detection and monitoring of fetal and neonatal brain development. Nowadays deep learning can help clinicians in performing these tasks achieving similar or even better results than classical approaches but much faster. V.Comte et al., 2023 presented a novel unsupervised segmentation method based on multi-atlas segmentation, which in fact, avoids the need for a large annotated dataset to train the DL network [1]. This work aims to modify this pipeline by incorporating a biomechanical model to serve as a constraint during CNN training. Biomechanical models that can describe brain deformation can be utilized for DL regularization purposes. Furthermore, this work also tries to access the local shear modulus of the fetal brain, an important measurement that otherwise is nearly inaccessible through other techniques. The best validation Dice score obtained in this work was 0.924 and outperforms some of the state-of-the-art registration pipelines that have been developed, where their validation Dice score was around 0.862 [2] and 0.858 [3]. Moreover, this work was used to obtain local shear and bulk modulus maps of the fetal bran where the cerebellum, the brainstem and the thalamus were found to be the stiffer regions in the brain. Additionally, the results obtained in this project indicate that the white matter exhibits slightly greater stiffness compared to the cortex.