Multi-class semantic segmentation of artifacts

Due to the presence of artifacts in digitized histopathology images, tissue regions that are important for diagnosis can be unclear or even completely unusable. As a result, artifacts may interfere with an accurate diagnosis by pathologists or lead to erroneous analyses by deep learning algorithms. By detecting commonly found artifacts, it is possible to improve the process of automated diagnosis. Images with too many artifacts can automatically be rejected by our quality control system, potentially saving many hours of manual inspection.

Included artifacts

The system detects the following artifact types:

1. Tissue folds
2. Ink
3. Dust
4. Air bubbles
5. Marker
6. Out-of-focus

Demo

An interactive demo is accessible via Grand-Challenge:

• URL: grand-challenge.org/artifact-segmentation

To be able to make use of the demo, please click on Request Access.

Dataset

The following slides were contained in the training data:

Tissue	Stain	Whole slide scanner	#Slides	#Annotations
Breast	HE CD3 CD45RO CD8	3DHistech Pannoramic 250 Flash II Hamamatsu NanoZoomer-XR C12000-01 Philips Ultrafast Scanner	47	1456
Bonemarrow	HE PAS	3DHistech P1000 scanners	8	71
Colon	CK20	3DHistech Pannoramic 250 Flash II	25	735
Kidney	HE PAS	Leica Aperio XT ScanScope 3DHistech P1000 scanners	9	211
Lymphoma	HE	3DHistech P1000 scanners	9	48
Pancreas	CD3 CK81 CKPAN CD8	3DHistech Pannoramic 250 Flash II 3DHistech P1000 scanners	10	451
Prostate	HE	Leica Aperio AT2 Hamamatsu C9600-12 3DHistech P1000 scanners	20	166
Ovary	HE P53 Ki67	3DHistech P1000 scanners	9	71
Miscellaneous	HE Ki67	3DHistech P1000 scanners	5	69

Inference

Prepare a configuration file (e.g., named config.yml):

# paths to checkpoints and output folder
artifact_network: /.../path/to/checkpoints/artifact_network.pth
quality_network: /.../path/to/checkpoints/quality_score_network.ckpt
output_folder: /.../path/to/output/directory
# architecure settings
architecture: deeplabv3plus
encoder_name: efficientnet-b2
encoder_weights: imagenet
use_batchnorm: true
pooling: avg
dropout: 0.4
spacing: 4.0
# torch settings
device: cuda
workers: 4
# tissue settings
minimum_region: 5000
minimum_hole: 1000
remove_mask: true
# artifact settings
overlap: 256
tile_size: 1024
tissue_spacing: 4.0
region_threshold: 10.0
hole_threshold: 200.0
num_classes: 7
# slides to process
input_path: /.../path/to/input/folder/*.tif
# tissue-background segmentation mask
mask_path: /.../path/to/tissue-background-masks/folder/{image}_tb_mask.tif

Run infer.py using the Docker image:

#!/usr/bin/env bash
cd "/home/user/source/artifact-segmentation/"
git pull
python3.7 /home/user/source/artifact-segmentation/src/infer.py \
  --config /.../path/to/config/config.yml

Training

Copy (and change paramaters of) the configuration file training.yml:

Run train.py using the Docker image:

#!/usr/bin/env bash
cd "/home/user/source/artifact-segmentation/"
git pull
python3.7 /home/user/source/artifact-segmentation/src/train.py \
  --config /.../path/to/config/training.yml

Requirements

Successfully tested using the following settings:

• --gpu-count="1"
• --require-gpu-mem="11G"
• --require-cpus="2"
• --require-mem="48G"