Top-k Bottom All but <i>σ</i> Loss Strategy for Medical Image Segmentation.

However, we argue that in the general medical image segmentation problem, the envisaged scenario is very likely: databases are small, although on an increasing trend, and contain both easy examples (used in medical teaching to explain the basics) and hard ones, which are needed in the later stages of learning (both for machines and medical students) and annotations at pixel level are sometimes noisy, the root cause being the limited time that experts allocate for labeling.

also in 63,971 other papers