Article Directory
Summary
We propose a new multi-class weighted loss function, such as the segmentation of cluttered cells. Our main motivation is that developmental biologists need to quantify and model the behavior of blood T cells, which may help us understand their regulatory mechanisms and ultimately help researchers seek an effective immunotherapy cancer treatment. Segmenting a single touch cell in a cluttered area is challenging because the shared boundary and the feature distribution on the cell foreground are similar, so it is difficult to distinguish pixels into appropriate classes. We propose two new weight maps to be applied to the weighted cross-entropy loss function, which takes into account class imbalance and cell geometry.
The binary training data label has been enhanced, so that the learning model can not only handle the foreground and background, but also the third touch category. The framework allows the use of U-Net