Volumetric segmentation of the corpus callosum: training a deep learning model on diffusion MRI

Can incomplete silver standard labels improve performance of DTI-based volumetric segmentation of the corpus callosum?

“…Those fibers perform the interconnections between the cortical and subcortical regions, which are necessary to the underlying neural networks, allowing complex activities. 1 Due to this, CC brings anatomical, functional, and clinical relevance. Some diseases can affect the CC directly, for example, Autism Spectrum Disorders (ASD), 2,3 Bipolar Disorder (BD) and suicidal ideation.…”

“…analyzing fractional anisotropy and corpus callosum volume). 1,9,13 CC manual segmentation done by experts is still the ground truth for many studies. 14 However, it requires training and is time-consuming, due to similarity and proximity of the fornix 15 and the cingulum.…”

“…Although segmenting CC on T1w images and registering to diffusion space is feasible, it introduces distortions through registration, has a high computational cost and execution time. 1 And although there is an increasing number of brain segmentation tools, some of them based on atlas, such as FSL, 17 some more recent based on Deep Learning, such as QuickNat, 18 none of them segments the CC. The only two tools available for subcortical structures segmentation that include the CC are: FreeSurfer, 19 which segments the CC only on seven slices; and MRICloud, 20 a cloud-based tool.…”

“…Among the existing semi-automated and automated segmentation methods working directly on DTI, 15 most of them segment the CC on the midsagittal slice such as the ones proposed by Freitas et al, 21 Niogi et al 22 Nazem-Zadeh et al, 23 Luis-Garcıa et al, 24 Cover et al 25 and Herrera et al 26 For volumetric CC segmentation on DTI, there are only five automated methods: Rittner et al 27 extended Freitas et al' 28 approach and achieved a hierarchical segmentation by applying the 3D watershed using automatically selected markers; Kong et al 29 proposed a method to automatically learn an adaptive distance metric, through a semi-supervised learning model based on graphs; Wang et al 30 introduced the gray information of multiple atlases and the prior information of target shapes into the Active Shape Model (ASM) and proposed the Multi-Atlas Active Shape Model (MA-ASM) approach for segmentation of 7 Region of Interest (ROI), including the CC; Rodrigues et al 1 developed a volumetric segmentation method for the CC using Convolutional Neural Networks (CNN) using as input diffusion maps, such as fractional anisotropy (FA), mean diffusivity (MD) and mode of anisotropy (MO), separately and combined; Rodrigues et al 31 developed a subsequent method based on 3D convolutions, presenting better performance in comparison to the 2D model.…”

“…Since improving further the performance of the volumetric segmentation proposed by Rodrigues et al 31 would require more annotated data, we decided to investigate here if the introduction of incomplete silver standard annotations from FreeSurfer would be a solution. This would avoid the cost of having complete 3D manual annotations of the CC on hundreds of subjects and allow us to move towards a more robust and general method, by adding images from diverse sources to our training dataset.…”