A Novel Framework for Coarse-Grained Semantic Segmentation of Whole-Slide Images

Abstract

Semantic segmentation of multi-gigapixel whole-slide images (WSI) is fundamental to computational pathology, as segmentation of different tissue types and layers is a prerequisite for several downstream histology image analysis, such as morphometric analysis, cancer grading, and survival. Both patch-based classification and pixel-wise segmentation have been used for these tasks, where patch-based classification outputs only one label per patch while pixel-wise segmentation is more accurate and precise but it requires a large number of pixel-wise precise annotated ground truth. In this paper, we propose coarse segmentation as a new middle ground to both techniques for leveraging more context without requiring pixel-level annotations. Our proposed coarse segmentation network is a convolutional neural network (CNN) with skip connections but does not contain any decoder and utilizes sparsely annotated images during training. It takes an input patch of size $M \times N$ and outputs a dense prediction map of size $m \times n$, which is coarser than pixel-wise segmentation methods but denser than patch-based classification methods. We compare our proposed method with its counterparts and demonstrate its superior performance for both pixel-based segmentation and patch-based classification tasks. In addition, we also compared the impact on performance of coarse-grained and pixel-wise semantic segmentation in downstream analysis tasks and showed coarse-grained semantic segmentation has no/marginal impact on the final results.