Weakly-supervised Semantic Segmentation for Histopathology Images

This repository contains an overview and short papers reviews of WSSS methods for histopathological image data. Occasionally, WSSS papers dealing exclusively with semantic segmentation on natural images (e.g. from PASCAL-VOC 2012 or ADE20K) will be included. I will update this list continuously during my current master thesis.

Progress on Literature Review

The ground-truth annotation type used for training is mentioned in the Annotation column for each paper. For WSSS, it generally falls into one of the following categories (in order of decreasing informativeness):

• Bounding box
• Scribble
• Point
• Image label

The LOD column roughly indicates the level of detail of the segmentation for each paper, e.g. whether a class is assigned to each pixel or only to each patch. This is a very vague information, since the pathology images have a different resolution for the different methods and the chosen patch size differs too. Therefore the LOD should only serve as a rough estimation for the time being.

WSSS for Histopathology Images

Conference	Title	Annotation	LOD	Notes	Official Code	Datasets
ICCV 2019	CAMEL: A Weakly Supervised Learning Framework for Histopathology Image Segmentation	image label (1280 x 1280)	pixel-level	yes	-	CAMELYON16, Colorectal Adenoma
MICCAI 2020	Weakly supervised multiple instance learning histopathological tumor segmentation	image label (100k x 100k)	patch-level (224 x 224)	yes	pytorch	TCGA, PatchCamelyon
2021	Data-efficient and weakly supervised computational pathology on whole-slide images	image label (100k x 100k)	patch-level (< 256 x 256)	coming soon	pytorch	CAMELYON16, CAMELYON17, TCGA, CPTAC
MICCAI 2021	Learning Whole-Slide Segmentation from Inexact and Incomplete Labels using Tissue Graphs	image labels + scribbles (11k x 11k and 3k x 3k)	superpixel-level	coming soon	pytorch	SICAPv2, UZH
MICCAI 2022	Transformer based multiple instance learning for weakly supervised histopathology image segmentation	image label (3000 x 3000)	pixel-level	yes	pytorch	Colon cancer
ICIGP 2022	HistoSegResT: A Weakly Supervised Learning Method for Histopathology Image Segmentation	image label (775 x 522)	pixel-level	coming soon	-	GlaS

WSSS for Natural Images

Conference	Title	Annotation	LOD	Notes	Offical Code	Datasets
ICCV 2017	Simple Does It: Weakly Supervised Instance and Semantic Segmentation	bounding box (multi-class)	pixel-level	coming soon	no	PASCAL VOC2012, VOC12+COCO

Histopathology Datasets

Below is an overview of histopathology datasets from various body regions and tumor types. These can be used either for training or evaluation of classification or segmentation algorithms. The columns have the following meaning:

• #Labels/Img: Number of labels per image (BG = background)
• #Classes: How many different classes an image can be contain (most datasets only differentiate between tumor and background)
• #Img: Total number of images in the dataset
• #PGT: Number of images which have a pixel-wise ground-truth annotation (some images may only have image-level labels)

Name	Type	#Labels/Img	#Classes	#Img	#PGT	Image Size	Resolution	Paper
CAMELYON16	Lymph node metastasis	1	1	400	400	100k x 100k	0.25 microns/px	yes
CAMELYON17	Lymph node metastasis	1	1	1000	500	100k x 100k	0.25 microns/px	yes
PatchCamelyon	Lymph node metastasis	1	1	327.680	0	96 x 96	0.25 microns/px	no
CoCaHis	Colon cancer	1 + BG	1 + BG	82	82	1037 x 1388	0.45 microns/px	yes