How to deal with overlapping segmentations for chest scan images

vision
1

Hi all,

I am working on a chest scan image classification model on pneumonia related conditions but I came to understand from clinicians that some of the conditions they are looking for actually overlaps one another on the chest scan image. Am aware that UNet is widely used for image segmentation but does it really solve the problem of overlapping images? If not what are other methods to solve this issue? I have done up a multilabel classification model but results aren’t great, as I would score highly for my F1 in one condition but some overlapping conditions will get a low F1 score.

One alternative I had was to split each conditions up into an individual binary classification model, but am trying to explore if there are other better alternatives. Am currently training my images on densenet121

2

You can do this by just re-interpreting the output of the u-net as a “multi-class” classification output (i.e. one that is put through sigmoid to get a probability for each class individually).

Best regards

Thomas