spinalcordtoolbox / disc-labeling-hourglass

Labeling of intervertebral discs using the Hourglass deep learning architecture.
GNU Lesser General Public License v2.1
1 stars 0 forks source link

Specifying the number of discs to detect using `--ndiscs` #1

Open NathanMolinier opened 1 year ago

NathanMolinier commented 1 year ago

Context

In this paper, the hourglass network was trained on multiple subject samples with a variable number of intervertebral discs : between 7 and 11 discs.

Issue

Therefore, when more discs are present in the image, the hourglass is unable to identify them, leading to false negative detections.

Parameter

The parameter --ndiscs used to specify the maximum number of class (number of discs) we want the hourglass to detect during training is therefore a key argument.

NathanMolinier commented 1 year ago

Creating a class for each disc seems to be the solution, however, this is only possible if each discs are represented in the training dataset