Learning The Optimal Threshold: A Novel Method to Segment Histopathology Images

In the last decade, a lot of segmentation techniques had been proposed. Most of them include using an encoder-decoder network, such as the UNet model, to predict the mask for a certain image. The issue with UNet models and fully convolutional networks, in general, is that they require a substantial amount of data. In this study, we propose a novel technique to segment nuclei from breast histology patches. Instead of learning to generate a mask from an input image, we simplify the problem by using a regression network that learns to predict the optimal threshold value. To find the best threshold value for a certain histology image, we introduce a similarity guided search algorithm that compares the embeddings of the ground truth mask with the embedding of all masks generated from different possible threshold values. The proposed network manages to segment the nuclei from a small histology dataset with high accuracy i.e. 90%.