newiz430 / cellsegmentation Goto Github PK

This is the new multi-stage auto-immunoquantitative analytical model we proposed.

Special thanks to Dr. Cheng @ShenghuaCheng for contributing to this work and WNLO for platform provision.

Taking immunohistochemistry-stained digital cell images as input, the model is merely supervised by positive cell counting labels and transforms whole-image (bag) level counting results into superpixel (instance) level classification results via the specifically designed adaptive top-k instance selection strategy.

• Stage 1: Image-wise regressive positive cell counter
• Stage 2: Superpixel-wise tile instance classifier
• Stage 3: Pixel-wise segmentation encoder-decoder network

Instance classifier provides us semantic information of positive cells. HSV channel separation and thresholding provide us fine-grained profile of positive cells.

Kappa = 0.9319, 4th in Lymphocyte Assessment Hackathon (LYSTO) Challenge. Leaderboard

We also tested our localization method in LYON19.

Visit LYSTO to get data.

• Add image data in ./data
• Train cell counter by python train_image.py
• Test your counter by python test_count.py
• Train tile classifier by python train_tile.py
• Test the classifier and get heatmaps by python test_tile.py
• Train segmentation network by python train_seg.py
• Test segmentation network and get masks by python test_seg.py --draw_masks
• Test segmentation network and get localization points by python test_seg.py --detect

and use arguments you like. You can find arguments list in the source code file.

... under construction ... stay tuned.

2021-2022 By Newiz