U-net Ensemble Model for Segmentation inHistopathology Images

In this work, a multi-scale U-net[1] fusion model is proposed for the automatic cancer detection and classification in whole-slide lung histopathology[2]. The model integrates two types of U-net structure, trained on different image scales and subsets, aiming to address the challenges posed by the significant variation in data presentation. Since lung histopathology images come in various sub-categories and appearances, the performance of an individual trained network is usually limited. We train a variety of networks by using multiple re-scaled images and different subsets of images, and finally ensemble the outputs of various networks. Smoothing and noise elimination are conducted using convolutional Conditional Random Fields (CRFs)[3]. The proposed model is validated on Automatic Cancer Detection and Classification in Wholeslide Lung Histopathology (ACDC@LungHP) challenge in ISBI2019. Our method achieves a dice coefficient of 0.7968, Which is ranked at the third place on the board.