Deep-BCR-Auto

This is the official implementation of Deep-BCR-Auto.

Tissue detection and patch cropping

1. Use any foreground segmentation method you like. Here, we use the CLAM's implementation. Clone and go to the /CLAM folder.
2. Basic run (changed sthresh to 20):

python create_patches_fp.py --source <> --save_dir results/TCGA --patch_size 448 --step_size 448 --seg

3. Tune segmentations. Inspect each mask, tune process_list_edited.csv and set processed to 1 for the mask you want to tune:

python create_patches_fp.py --source <> --save_dir results/TCGA --patch_size 448 --step_size 448 --seg --process_list process_list_edited.csv

4. Get tissue patches. Set all processed to 1:

python create_patches_fp.py --source <> --save_dir results/TCGA --patch_size 448 --step_size 448 --seg --process_list process_list_edited.csv --patch --stitch

Tumor bulk detection

1. GO TO /tumorbulk
2. Download pretrained weights from here .
   Copy 'ctranspath.pth' to /tumorbulk/TransPath/ and copy 'model_best.pth.tar' to /tumorbulk/
3. generate tumor bulk masks

python tumorbulk.py --datadir <> --ptsdir ../CLAM/results/TCGA/patches --savedir ./results --code TCGA

3. create a dataset excel based on those having masks. ('../tcga_brca_bulkdetected.csv')
4. crop patches:

python pts_extraction.py -p 896 -s 896 -l 0 --code TCGA --datadf ../tcga_brca_bulkdetected.csv

Main args:

• -p : patch size

• -s : stride

• -l : patch extraction at which level (i.e., openslide downsample level)

• --code : experiment code

• --datadf : data spread sheet (i.e., dataset excel)

• **NOTE: pts are now saved in (x, y) format for consistency

Deep-BCR

1. go to ./casii/TransPath. Copy 'ctranspath.pth' to /casii/TransPath/ (see the google link from previous step).
2. encode patches:

python get_features_CTransPath_stainnorm.py --psize 896 --level 0 --datadf ../../tcga_brca_bulkdetected.csv --save ../data/ctpnormembedding --stainnorm

Main args:

• --psize : patch size
• --stride : stride
• --level : openslide level
• --datadf : dataset spread sheet. listed all the slide that you are using.
• --stainnorm : perform stain norm. store true, default: False
• --save: save dir

3. *(optional) encoding using resnet50:

python get_features_resnet_stainnorm.py --psize 896 --level 0 --datadf ../../tcga_brca_bulkdetected.csv --save ../data/resnormembedding --stainnorm 

3. build keyset:

python keyset_lrn.py --datadf '../tcga_brca_bulkdetected.csv' --featdir './data/ctpnormembedding/l0p896s896' --task fivefold -t 100 --psize 896

Main args:

• --featdir : feature dir
• --task : task name
• -t : maximum # of keys per slide
• --cur : run cur function. store true.
• --extract : run keyset extraction function (must be after cur). store true.
• --psize : patch size
• --encoder : encoder name. default: ctpnorm

4. run casii:

python train.py --arch CASii_MB --data threefold --code ctpnormTCGAodx_patience5_stopep10_ws_lr1e4 --psize 896 --nfold 3 --weighted-sample --patience 5 --stop_epoch 10 --lr 1e-4

Main args:

• --arch : model arch
• --data : dataset name in mydatasets
• --ctp_recurnot : exp code
• --psize : patch size
• --weighted_sample : store-true, default: False
• --nfold : num of folds, default: 5

Pretrained weights can be find at here in folders 42, 43, and 44. 42 is the best performed fold.
Our splits file can be find at the same link in folder splits.

5. test

python eval.py --arch CASii_MB --data threefold --code ctpnormTCGAodx_patience5_stopep10_ws_lr1e4 --psize 896 --nfold 3

Main args: Main args:

• --arch : model arch
• --data : dataset name in mydatasets
• --ctp_recurnot : exp code
• --psize : patch size