Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Label merge-and-split: A graph-colouring approach for memory-efficient brain parcellation

Aaron Kujawa, Reuben Dorent, Sebastien Ourselin, Tom Vercauteren

TL;DR

This work introduces label merge-and-split, a method that first greatly reduces the effective number of labels required for learning-based whole brain parcellation and then recovers original labels and achieves segmentation accuracy comparable to the baseline method without label merging and splitting.

Abstract

Whole brain parcellation requires inferring hundreds of segmentation labels in large image volumes and thus presents significant practical challenges for deep learning approaches. We introduce label merge-and-split, a method that first greatly reduces the effective number of labels required for learning-based whole brain parcellation and then recovers original labels. Using a greedy graph colouring algorithm, our method automatically groups and merges multiple spatially separate labels prior to model training and inference. The merged labels may be semantically unrelated. A deep learning model is trained to predict merged labels. At inference time, original labels are restored using atlas-based influence regions. In our experiments, the proposed approach reduces the number of labels by up to 68% while achieving segmentation accuracy comparable to the baseline method without label merging and splitting. Moreover, model training and inference times as well as GPU memory requirements were reduced significantly. The proposed method can be applied to all semantic segmentation tasks with a large number of spatially separate classes within an atlas-based prior.

Label merge-and-split: A graph-colouring approach for memory-efficient brain parcellation

TL;DR

This work introduces label merge-and-split, a method that first greatly reduces the effective number of labels required for learning-based whole brain parcellation and then recovers original labels and achieves segmentation accuracy comparable to the baseline method without label merging and splitting.

Abstract

Whole brain parcellation requires inferring hundreds of segmentation labels in large image volumes and thus presents significant practical challenges for deep learning approaches. We introduce label merge-and-split, a method that first greatly reduces the effective number of labels required for learning-based whole brain parcellation and then recovers original labels. Using a greedy graph colouring algorithm, our method automatically groups and merges multiple spatially separate labels prior to model training and inference. The merged labels may be semantically unrelated. A deep learning model is trained to predict merged labels. At inference time, original labels are restored using atlas-based influence regions. In our experiments, the proposed approach reduces the number of labels by up to 68% while achieving segmentation accuracy comparable to the baseline method without label merging and splitting. Moreover, model training and inference times as well as GPU memory requirements were reduced significantly. The proposed method can be applied to all semantic segmentation tasks with a large number of spatially separate classes within an atlas-based prior.
Paper Structure (13 sections, 6 equations, 4 figures, 1 table)

This paper contains 13 sections, 6 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Preprocessing.
  4. Merging of labels.
  5. Convolutional neural network.
  6. Splitting of labels.
  7. Evaluation methodology
  8. Data sets.
  9. Model training and testing.
  10. Baseline and evaluation metric.
  11. Results
  12. Discussion and conclusion
  13. Disclosure of Interests.

Figures (4)

  • Figure 1: Method overview illustrated on a toy example. a) Training. Original labels in MNI space are automatically grouped using a greedy graph colouring algorithm and merged. Toy example: {blue, red, purple} $\rightarrow$ green; {gray, black} $\rightarrow$ brown; and {yellow} $\rightarrow$ yellow. The CNN is trained on merged labels. b) Inference. The CNN predicts merged labels. In a post-processing step, original labels are recovered using one influence map for every set of merged labels.
  • Figure 2: Label merging. The distance between original labels is calculated from (MNI aligned) training set label volumes. A distance threshold and is applied to obtain an adjacency matrix (yellow: adjacent, blue: non-adjacent). Similarly, an average label volume ratio threshold is applied (not shown here). The adjacency matrix corresponds to an undirected connected graph, to which a greedy graph colouring algorithm is applied. Labels of the same colour are grouped, forming a new merged label.
  • Figure 3: Label splitting. A fuzzy prior is created from all (MNI aligned) training set label volumes. Subsequently, a Euclidean Distance Transform is applied for each label. One influence region map is created for each merged label. In this figure, the image on the right is the influence region map for the merged label that was merged from the right lateral ventricle label and 6 other original labels.
  • Figure 4: Effect of the label merge-and-split approach on the quality of predictions. Dice Similarity Coefficients (DSC) are calculated with respect to the GIF pseudo-groundtruth (AOMIC and IXI testing sets) or manual groundtruth (Mindboggle42). Each data point is the average DSC of all labels. Boxplots with DSCs of all individual labels are included in the supplementary material. Green triangles represent the mean. Asterisks indicate statistical significance according to a Wilcoxon test (paired samples) with Bonferroni correction where *, **, ***, and **** indicate p-values below 0.05, 0.01, 0.001, 0.0001, respectively, and ns (not significant) p-values above 0.05.