By: CVIP on Nov. 22, 2021, 6:46 a.m.
When I tested the simulated bboxes that you provide for the clean CXR images. I found some of the bboxes are located at the non-lung regions. Is this setting reasonable? Hope you can give me some help.
By: ecemsogancioglu on Nov. 22, 2021, 9:40 a.m.
Hi,
Thanks for your question, could you mention a filename where this happens as an example just to doublecheck?
By: ecemsogancioglu on Nov. 23, 2021, 11:09 a.m.
Hi,
The example images that you shared are still (expected to be ) in the lung region. Because there is still some lung in the areas that do not look dark, such as behind the heart.
We used a lung segmentation algorithm to select a region where we could pick random nodule locations as in the example link So we select the area that is denoted with the blue line, and pick a random nodule locations within this area.
I think it is not a perfect estimation of the lung area, because it is not easy to estimate the full volume from a 2D CXR. But it should still be good enough. There is a chance that this can create few bounding boxes that are not in the lung region, but we expect that will not harm the algorithm performance.
