Infant head and brain segmentation from magnetic resonance images using fusion-based deep learning strategies

Abstract Magnetic resonance (MR) imaging is widely used for assessing infant head and brain development and for diagnosing pathologies. The main goal of this work is the development of a segmentation framework to create patient-specific head and brain anatomical models from MR images for clinical evaluation. The proposed strategy consists of a fusion-based Deep Learning (DL) approach that combines the information of different image sequences within the MR acquisition protocol, including the axial T1w, sagittal T1w, and coronal T1w after contrast. These image sequences are used as input for different fusion encoder–decoder network architectures based on the well-established U-Net framework. Specifically, three different fusion strategies are proposed and evaluated, namely early, intermediate, and late fusion. In the early fusion approach, the images are integrated at the beginning of the encoder–decoder architecture. In the intermediate fusion strategy, each image sequence is processed by an independent encoder, and the resulting feature maps are then jointly processed by a single decoder. In the late fusion method, each image is individually processed by an encoder–decoder, and the resulting feature maps are then combined to generate the final segmentations. A clinical in-house dataset consisting of 19 MR scans was used and divided into training, validation, and testing sets, with 3 MR scans defined as a fixed validation set. For the remaining 16 MR scans, a cross-validation approach was adopted to assess the performance of the methods. The training and testing processes were carried out with a split ratio of 75% for the training set and 25% for the testing set. The results show that the early and intermediate fusion methodologies presented the better performance (Dice coefficient of 97.6 ± 1.5% and 97.3 ± 1.8% for the head and Dice of 94.5 ± 1.7% and 94.8 ± 1.8% for the brain, respectively), whereas the late fusion method generated slightly worst results (Dice of 95.5 ± 4.4% and 93.8 ± 3.1% for the head and brain, respectively). Nevertheless, the volumetric analysis showed that no statistically significant differences were found between the volumes of the models generated by all the segmentation strategies and the ground truths. Overall, the proposed frameworks demonstrate accurate segmentation results and prove to be feasible for anatomical model analysis in clinical practice.