Deep Learning-based Histopathological Segmentation for Whole Slide Images of Colorectal Cancer in a Compressed Domain with Wavelet Weighted Ensemble

Abstract

Analyses of histopathological images are the key steps for cancer diagnosis and progression. However, their accuracies are varied by pathologists relying on their expertise and other factors. Further, the annotation of the pathological images is a time-consuming process. Thus, automatic pattern recognition using deep learning techniques becomes increasingly important. To analyze a large scale of histopathological images, decimation or tile extraction is an essential process in deep learning due to limited system memory. However, not only the conventional decimation may lose high-frequency information but also the tile extraction without decimation may decrease the size of the region of interest. To overcome these limitations, here, we propose an image segmentation approach with compression, which is typically issued in JPEG 2000. We adapted U-net++ as a semantic segmentation model to train compressed tiles. After inference for each tile, a whole prediction image is reconstructed by wavelet weighted ensemble based on inverse wavelet transform. The training and validation are performed using hematoxylin and eosin-stained whole slide images of 391 colorectal biopsy specimens which are pathologically confirmed in the Yeouido St. Mary’s Hospital, the Catholic University of Korea. For the test dataset, the average Dice score is 0.852 ± 0.089 and the pixel accuracy is 0.962 ± 0.027. We believe that our approach has a great potential to perform fast and accurate diagnosis in pathology and that it can be also applied to other fields where large images must be processed.