About prediction and dataset.json

[mclvnot]

Hello,

After 5 fold training and consolidated, I wanted to predict dataset with _nndet_predict 006 RetinaUNetV001_D3V0013d --fold -1 for consolidated results. However, I faced Found different values for fold, will overwrite 0 with -1 warning. It is normal for prediction with consolidated results that exists in consolidated folder. Why do I get this warning?

Also, for multiclass and labelled dataset, what do i write "target class" and "labels" in dataset.json? or does nndetect support multiclass?

Best,

[mibaumgartner]

Hi,

the warning is indeed normal and will be fixed/removed in the future, thanks for pointing it out.

"labels" should contain all the classes from your dataset, otherwise there is now way how nnDetection would know what classes are present. "target_class" is only needed for a specific use case with patient level evaluation, e.g. for a dataset with benign and malignant lesions we are usually only interested in the patient level performance of the malignant class, so target class could be set to that class. The default object level evaluation (with the COCO metric - mAP) is always performed in a multi class fashion (i.e. evaluating each class separately and aggregating afterwards). The current FROC implementation handles multi-class evaluation differently and will be replaced/complemented with a different implementation in the future.

[mclvnot]

Thank you!

Could we convert prediction results (zmin, zmax, xmin, xmax, ymin, ymax) to 2D predictions?

Best,

[mibaumgartner]

You could interpret them slice by slice but it is still different from a usual 2D prediction which was derived slice by slice since it represents the maximal extend of the object along the z axis.

[mclvnot]

Also, I want to ask about evaluation metrics. In test results, I take care of _AP_IoU_0.10_MaxDet100. Intersection over union threshold is 0,1 but is confidence score threshold is 0?

Best,

[mibaumgartner]

The evaluation does not apply any confidence threshold but there is a maximum of 100 predictions per image. There might a confidence threshold in the model though, since sometimes it can boost performance before running the weighted box clustering

[github-actions[bot]]

This issue is stale because it has been open for 30 days with no activity.

[github-actions[bot]]

This issue was closed because it has been inactive for 14 days since being marked as stale.