Backpropagated

In my previous post I detailed methods for conducting a binary image segmentation of the corpus callosum. I used three different image modalities (T1-weighted, T2-weighted and generalized fractional anisotropy) and a simple k-Nearest Neighbors (kNN) model from sklearn. We also covered some basic normalization that is important for kNN models. Today’s goal is to demonstrate how some knowledge of your dataset can help you derive informative features. In our dataset, we need to deal with some misclassified tissue from my last post on the corpus callosum. [Read More]

Atrophy of the corpus callosum is an established quantitative biomarker in several neurodegenerative diseases. For instance, in multiple sclerosis atrophy of the corpus callosum is associated with whole-brain atrophy (Klawiter et al., 2014), increased cognitive disability (Llufriu et al., 2012) and altered interhemispheric functional connectivity (Tobyne et al., 2016). Several callosal segmentation solutions exist, and the subject been published on extensively; however these tools often rely upon spatial realignment to an atlas, use complex shape-deformation based algorithms requiring adequate compute power or have not been publicly released. [Read More]