Releases: KatherLab/STAMP
Releases · KatherLab/STAMP
2.4.0
- Added support for regression (MAE loss) and Cox-based survival analysis for tile-, slide-, and patient-level tasks.
- Added evaluation metrics:
- Regression: R², Pearson r, Pearson p, MAE, RMSE
- Survival: Kaplan–Meier curve (optional: low/high risk using a median cut-off computed from train predictions), log-rank p-value, C-index
- Added probability heatmaps for regression and survival tasks.
- Training now supports Attention (ViT), TransMIL, and MLP backbones.
- All backbones support tile-level modeling; slide/patient-level modeling is supported via MLP
- Automatic model class selection based on feature type and task.
2.3.0
- Modeling and statistics support for patient-level features.
- MCP support for all STAMP features.
- Fixed heatmaps and top tiles
- Overlayed heatmaps on thumbnail.
- Easier installation steps.
- Restored Learning Rate Scheduler and original STAMPv1 hyperparams.
- Advanced params section in config file for easier tuning without modifying the code.
- Added PRISM.
2.2.0
- Slide and patient level encoding support with:
- COBRA
- EAGLE
- CHIEF
- TITAN
- GIGAPATH
- MADELEINE
- New tile-level feature extractors:
- Virchow
- mSTAR
- PLIP
- MUSK
- Fixed high memory usage disabling attention masking on validation and testing.
- Option to add hash to output folder.
- Option to change cached tiles to png.
- Updated docs.
- Minor bug fixes.
2.1.0
v2.0.0
- Better and stricter config parsing
- Significantly faster preprocessing
- No more dependency on fastai
- Lots of code cleanup, refactoring, documentation and tests
2.0.0.dev14
Add tests for heatmaps
2.0.0-dev8
Implement masking
STAMP v1.1.1
- Deploy now happens on GPU when available
- Minor heatmap bug fix for resizing a PIL image using a tuple instead of an array
STAMP v1.1.0
- Replace MultiHeadAttention with PyTorch nn.MultiHeadAttention which uses FlashAttentionV2, thus requiring less memory while being faster;
- Enabled usage of TensorFloat32 on CUDA devices, resulting in faster performance;
- Changed optimizer from Adam to AdamW.
STAMP v1.0.3
- Integration of UNI feature extractor
- Move download of UNI model weights to
stamp setup - Add process ID to logfile name to avoid two slurm processes writing to same logfile
- Ignore missing normalization template / weights if not required by config
- Remove lock files older than 20 seconds before beginning preprocessing
- Fix of error in heatmaps script
- Set
norm: falsein the default config - Set version to 1.0.3
- Update installation instructions in README