Change probability threshold in prediction

[agatadefr]

Hello,

I would like to change threshold in prediction to detect more lesions ?
Where could I do that in the code ?

I would appreciate for any suggestion.

[FabianIsensee]

Hi, you would need to change the way the segmentation is obtained from the softmax probabilities.

Take a look at

nnUNet/nnunet/inference/predict.py

Line 131 in f5e72b1

def predict_cases(model, list_of_lists, output_filenames, folds, save_npz, num_threads_preprocessing,

In here we use multi processing to export the segmentations. Exporting first does resampling, then argmax and finally the saving of the result:

nnUNet/nnunet/inference/predict.py

Line 262 in f5e72b1

results.append(pool.starmap_async(save_segmentation_nifti_from_softmax,

So if you want to change the way this is done you need to change predict_cases and also save_segmentation_nifti_from_softmax (the latter is here:

nnUNet/nnunet/inference/segmentation_export.py

Line 27 in f5e72b1

def save_segmentation_nifti_from_softmax(segmentation_softmax: Union[str, np.ndarray], out_fname: str,

)