usage of masks in training about chestx-ray-14 HOT 1 CLOSED

Hi, Deep Learning method is a very powerful feature learning machine, so it can be very easily learn irrelevant features for the task at hand. So by training on the cropped X-ray, you explicitly force the model to learn more discriminative features on the lung region, which is where the thoracic disease occurs and visible in the chest X-ray.

Another idea is to use the mask generated by the model as pseudo-label and train segmentation and classification jointly with segmentation as auxiliary task, the combine loss can be L_combine = L_classification + 0.1*L_segmentation. This can squeeze out a little extra point. I haven't test it yet, but you can try.

Hope this answer your question.

from chestx-ray-14.