About the performance of BCSS dataset

[o0t1ng0o]

Hi @pengsl-lab,

Thank you for sharing your code. I have tried to reimplement the performance of BCSS dataset.

python3 train.py --dataset BCSS --network DHUnet \
--cfg configs/DHUnet_224.yaml --root_path data \
--max_epochs 150 --output_dir model-BCSS/DHUnet \
--img_size 224 --base_lr 0.005 --batch_size 24

But I found that I can only get this performance : mean dice 0.562457 yc 0.461954 acc 0.774036 And the results in your paper is as follows:

Is there any problem about my reimplementation?

Thank you in advance!

[pengsl-lab]

According to BCSS reference, we tried the Class grouping and Train-test splits procedures. 1、Not all classes are used, mapping pixels into five region classes: tumor, stroma, inflammatory infiltration, necrosis and others. (This may be the reason for the dissatisfied effect.) 2、Use an unseen testing set to report performance: OL, LL, E2, EW, GM, and S3. Our preprocessed data and trained model are publicly available at [https://drive.google.com/drive/folders/1cEHr1YPE3fuJ0AKlJtWHOZ_vLKMcdQPa?usp=sharing]. Please feel free to contact me if you have any further questions.

[guascy666]

i have the same problem with the above one。i have used your trained model but the acc result is only about 0.85。