Let’s consider a 2-class / binary segmentation problem where
c=0 for background (healthy tissues) and
c=1 for foreground (diseased tissues).
A trained U-Net model provides pixel-wise prediction either
1 for the foreground class
c=1 only. The prediction for
c=0 is obtained by inverting the model's output. Here is an example for a
3×3 image. The ground truth and predicted masks are given for both classes. Several performance matrices i.e., accuracy, precision, recall, dice, and iou are computed.
What's the standard practice in research community for publishing results in journals/conferences? Should the metrics be reported for c=1 (option 1) or mean of the both class (option 2)?
IoU(c=1) = 0.750i.e., only diseased tissues are important
mIoU = (IoU(c=0) + IoU(c=1)) / 2 = (0.833 + 0.750) / 2 = 0.792i.e., both classes are important and we can't afford mislabeling a healthy tissue as a diseased one or vice versa.
I will highly appreciate if you share some resource as reference.