This is a further version of the original unil_chuv method submitted for task 1 at MICCAI 2020.

Changes relative to first submission:

  • increased overlapping in the sliding window approach (before it was 0, now it’s 25%). This was probably the thing which had the biggest impact on sensitivity
  • resampled all volumes to median voxel spacing
  • removed FP smaller (in mm^3) than the smallest of all ADAM aneurysms
  • increased the number of training 3D patches extracted for each subject, from 30 to 40
  • Decreased the UNet threshold from 0.7 to 0.4
  • used bias-field corrected volumes instead of original
  • checked patients where registration was not correct: if it was wrong, I did not take the distances to the landmark points int account during the sliding window