Bottleneck Layer Embedding Differences in Semantic Segmentation of Multi-Source Cardiac MRI

Recent research has shown advances in the analysis of cardiovascular MRI imagesusing deep learning. However, two problems are apparent: How to measure the quality of theresult of semantic segmentations and how to expose dependencies on the actual MRI apparatusused in obtaining the image data sets. The proposed method is based on traditional evaluationsat the pixel level. Admittedly, it would be convenient to judge incoming samples on their fa-miliarity in relation to the training data. This would allow for filtering out inadequate samples.In order to solve this conveniently, it is proposed to compare incoming samples to prototypicalcentroid vectors in an embedding (sub space), by using dimensionality reduction. MRI imagesused for this experiment are fed through a fully connected network model trained on short-axisMRI’s of left ventricles. The machine learning model was tested using two different data setscollected from two different MRI devices, one generating the UK Biobank data and another,UMCG’s data. The raw MRI’s and the resulting segmentations are used for investigating theproblem of finding a reliable comparison method for judging whether an input sample meetsthe expectations that are represented by the statistics of the training data. To achieve this, adimensionally reduced representation of the data is calculated with which centroids can be com-puted for classes. Both are then used as dimensionally reduced representations of the data andaveraged to represent the centroid of their embedding. An optimal measurement is discoveredamong three standard distance calculations (SAD, SSD and mean correlation), that is, SAD.This was the best measurement of similarity in raw MRIs (non-segmented) as well as serving asa predictor of segmentation quality, as verified by the Dice metric.