DearCaat
commented
3 hours ago First, I apologize for any inconvenience caused by the survival reproduction-related issues. I'll try my best to answer your questions:
- The survival-related experiments in the paper were completed by co-authors, using different environments and features from my uploaded version, hence the discrepancies
- I later reproduced related experiments (R50 and UNI features, excluding PLIP features) in my uploaded environment. I can provide my reproduction results and all parameters used for reference.
- Since the code used in subsequent reproduction is not entirely consistent with the GitHub code, there might be some deviations. Please let me know if the deviations are significant, and I can provide the latest code and continuous support.
| Reproduction results and parameters (5-fold cross-validation): | Survival | BLCA | LUAD | LUSC | BLCA (UNI) | LUAD (UNI) | LUSC (UNI) |
|---|---|---|---|---|---|---|---|
| RRT-MIL | 61+2.11 | 64.01+4.39 | 60.99+6.93 | 61.54+4.32 | 66.38+2.58 | 62.63+4.18 | |
epeg_k=21,crmsa_k=1, region_num=20 |
epeg_k=17,crmsa_k=3, region_num=16 |
epeg_k=21,crmsa_k=1, region_num=20 |
epeg_k=9,crmsa_k=5, region_num=20 |
epeg_k=9,crmsa_k=5, region_num=16 |
epeg_k=9,crmsa_k=3, region_num=20 |
qingze-bai
It's an exciting job, but I'm nowhere near the performance of the paper's implementation when I reproduce the Survival Prediction task, roughly 3 points less for each task? I followed the commands provided by github exactly and performed the grid search, is there any other parameter that needs to be changed?