/elsa

evaluation of lesion segmentation algorithms for T1w stroke data

Primary LanguageHTML

Evaluation of Lesion Segmentation Algorithms

This github repository contains the data and analysis files we used to evaluate automated segmentations against one another (ALI, lesion_gnb, and LINDA).

/Image_Metrics

  • this directory contains the csv file with the image metrics we analyzed ("Image_Metrics_132subj.csv"dice, assd, hausdorff's distance, precision, recall); and the R markdown file that performed Friedman's Tests on each metric, followed by pairwise comparisons.
  • Figures 2 and 3 were generated from this manuscript
  • 132 subjects are included (of the original 181), as cases with 1. no detected lesioned voxels (39) and 2. cases in which all 3 algorithms had 0 overlap between automated mask and manual segmentation (10) were removed.

/Lesion_Volumes

  • this directory contains the analyses and data used to compare lesion volumes of automated masks and manual segmentations.
  • The R markdown file contains correlations between lesion volumes for each automated method, with and without adjusting for outliers using cook's distance
  • again, 132 subjects were included in this analysis

/Lesion_Characteristics

  • this directory contains data and analyses for the Fisher's exact tests used to analyze whether performance on automated methods varied by lesion size or stroke territory

  • the csv file contains data for all 181 subjects

     key:
     strokeType: 1= cortical, 2=subcortical, 3=brainstem, 4=cerebellar
     LesHem: 0=left hemisphere stroke, 1=right hemisphere stroke
     LesVol_Categorical: 1=small, 2=medium, 3=large
    

/Minimum_Distance

  • this directory contains the R script used to create Figure 4 in the manuscript.
  • We provide minimum distance as descriptive data; no statistical analysis is performed as the number of lesions that misclassified lesions (i.e., had 0 overlap between automated mask and manual segmentations) were not equal between methods