Skip to content

DL-model for multi-class tissue segmentation in colorectal cancer H&E slides, developed as part of the SemiCOL2023 Challenge.

License

Notifications You must be signed in to change notification settings

lely475/CTPLab_SemiCOL2023

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Multi-task learning for tissue segmentation and tumor detection in colorectal cancer histology slides

Lydia A. Schoenpflug, Maxime W. Lafarge, Anja L. Frei, Viktor H. Koelzer, Department of Pathology and Molecular Pathology, University Hospital and University of Zurich

The semi-supervised learning for CRC detection (SemiCOL) challenge 2023 provides partially annotated data to encourage the development of automated solutions for tissue segmentation and tumor detection. We propose a U-Net based multi-task model combined with channel-wise and image-statistics-based color augmentations, as well as test-time augmentation (see Figure 1), as a candidate solution to the SemiCOL challenge. Our approach achieved a multi-task Dice score of .8655 (Arm 1) and .8515 (Arm 2) for tissue segmentation and AUROC of .9725 (Arm 1) and 0.9750 (Arm 2) for tumor detection on the challenge validation set.

inference_pipeline3 Figure 1: Proposed approach to tissue segmentation and tumor detection: We train a multi-task U-Net-based model for segmentation and classification, but only use the segmentation branch during inference. The tumor detection score is computed based on the amount of predicted class pixels.

Setup

pip install -r requirements.txt

Generate tiles for training

Add location of the data folder (parent folder of 01_MANUAL and 02_WEAK) and desired output folder, where the generated tiles will be stored. Approximate required storage for the tiles is 175 GB.

python3 src/run_data_generation.py --o path/to/output/dir --data path/to/data/dir

Train model

Train model on tiled challenge training set, tensorboard outputs and ckpts are logged to the output folder. Arm configuration can be set as 1 or 2, to train with image-statistics-based augmentation using references from SemiCOL only or SemiCOL and MIDOG references.

python3 src/run_train_multi_task_learner.py --o path/to/output/dir --arm 1 or 2

Checkpoints

The following checkpoints are provided in the ckpts folder:

--arm1.ckpt: Model trained on the challenge training set, using references from the SemiCOL challenge for image-statistics-based augmentation.
--arm2.ckpt: Model trained on the challenge training set, using references from the SemiCOL and MIDOG challenge for image-statistics-based augmentation.

About

DL-model for multi-class tissue segmentation in colorectal cancer H&E slides, developed as part of the SemiCOL2023 Challenge.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages