Regularization Strategy for Multi-organ Nucleus Segmentation with Localizable Features

Abstract

Applying color normalization on H&E images is a famous protocol in digital pathology. Recently, the CutMix technique has a strong ability to generalize the classification models. In this paper, we propose the modified CutMix for segmentation tasks. We apply it to the MoNuSeg dataset. The U-Net with a MobileNet backbone is used for training and inferencing. Moreover, we compare it with the traditional color normalization. The results show that our modified CutMix outperformed color normalization with the +0.179 AJI score. It achieved the IoU score faster and got a higher AP for every IoU threshold.