Ensemble of 3D CNNs for Brain MRI Segmentation

Generating pseudo computed tomography (CT) images from magnetic resonance images (MRI) is an essential task for a variety of medical imaging applications. This task is performed by segmenting the brain image into multiple tissue classes: bone, soft tissue, and air where bone segmentation is the most challenging part. The aim of this work is to propose an ensemble of excitation-based CNNs for enhancing the bone segmentation in brain T1-weighted MRI. In order to create diversity, multiple 3D excitation-based CNNs which employ different combinations of activation and loss functions are aggregated to build an ensemble model that provides reliable segmentation results. Using 3D clinical MR and CT images of fifty patients, the proposed approach is evaluated by performing a comparison between segmented MRI and segmented CT. This work is also validated with the baseline approach and other methods reported in the literature. The ensemble model with three CNNs shows superior segmentation results compared to a single CNN model by considering the three tissue classes where the DSC is improved from 0.618 to 0.657, from 0.907 to 0.914, and from 0.965 to 0.962 for bone, soft tissue, and air classes, respectively. The presented ensemble model reveals its potential to boost the segmentation results compared to the segmentation results obtained by a single CNN.