Skip to main content

Advertisement

Table 4 Analysis of 900 CNN segmentations with available ground truth

From: Automated quality control in image segmentation: application to the UK Biobank cardiovascular magnetic resonance imaging study

Class Acc. TPR FPR MAE
     DSC MSD RMS HD
      mm mm mm
LVC 0.998 1.000 0.000 0.082 0.386 0.442 1.344
  1.000 1.000 0.000     
LVM 0.051 1.000 0.001 0.268 0.510 0.547 2.127
  1.000 1.000 0.000     
RVC 0.901 1.000 0.033 0.146 0.588 0.656 2.086
  0.997 0.997 0.000     
Av. 0.650 1.000 0.011 0.165 0.495 0.548 1.852
  0.999 0.999 0.000     
WH 0.998 1.000 0.000 0.089 0.460 0.509 1.698
  1.000 1.000 0.000     
  1. CNN segmentations as in Bai et al. [4]. Manual contours were available through Biobank Application 2964. Classes are LV Cavity (LVC), LV Myocardium (LVM), RV Cavity (RVC), an average over the classes (Av.) and a binary segmentation of the whole heart (WH). First row for each class shows the binary classification accuracy for ‘poor’ and ‘good’ segmentations in the Dice Similarity Coefficient (DSC) ranges [0.0 0.7) and [0.7 1.0] respectively. Second row for each class shows the binary classification accuracy for ‘poor’ and ‘good’ segmentations in the Mean Surface Distance (MSD) ranges [>2.0mm] and [0.0mm 2.0mm] respectively. True-positive and false-positive rates are also shown. We report mean absolute errors (MAE) on the predictions of DSC and additional surface-distance metrics: root-mean-squared surface distance (RMS) and Hausdorff distance (HD)