Would blank images affect segmentation output?

Currently, I am working on a project which works with MR images. Each input file consists of 3 dimensions: X, Y and temporal(time).

Lot of “slices” in the 3rd dimension are blank. Like this image (ignore the black backround).

The corresponding training labels (segmentation masks) are also blank. Will this affect the accuracy of my model?

What I understand is that you have a 2D view of one part of a patient and you are seeing how it moves over time, am I right?

Are you using a 3D CNN or a 2D CNN? For a 2D segmentation network, black slices should not be a problem, I would remove them for training and it should be fine at inference time.
If you are using a 3D CNN, the black slices will work similarly to augmentations like cutout or hide-and-seek where part of the image is removed. So, it is ok, it will just make it harder to train. Well, if the number of black slices is too large compared to the number of normal slices then the network may try to always get the background class. One workaround is to weigh more the segmented areas in the loss function.

If you have a lot of data and the rate of black slices is not too high I would train and see what happens. If you have a decent amount but not too many images, you can actually remove the black slices and try to train the model using only the previous detection.

Hope this helps

1 Like

I am using a 2D CNN. And thanks a lot for the reply!