README.md

---
license: cc-by-4.0
task_categories:
  - image-classification
tags:
  - medical
  - pathology
  - whole-slide-imaging
  - cancer-detection
---

# CPath Preprocessed Patch for E2E Training

<!-- Provide a quick summary of the dataset. -->

This dataset is used for training E2E CPath model, presented in [Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology](https://huggingface.co/papers/2506.02408), NeurIPS 2025.

**Code:** [https://github.com/DearCaat/E2E-WSI-ABMILX](https://github.com/DearCaat/E2E-WSI-ABMILX)

## Sample Usage

This dataset provides preprocessed whole-slide image (WSI) patches in LMDB format. Below is an example of how to load data from an LMDB dataset.

```python
import lmdb
import torch
import pickle
from datasets.utils import imfrombytes # Ensure this utility function is correctly referenced

slide_name = "xxxx"  # Example slide name
path_to_lmdb = "YOUR_PATH_TO_LMDB_FILE" # e.g., "/path/to/my_dataset_256_level0.lmdb"

# Open LMDB dataset
env = lmdb.open(path_to_lmdb, subdir=False, readonly=True, lock=False, 
                readahead=False, meminit=False, map_size=100 * (1024**3))

with env.begin(write=False) as txn:
    # Get patch count for the slide
    pn_dict = pickle.loads(txn.get(b'__pn__'))
    if slide_name not in pn_dict:
        raise ValueError(f"Slide ID {slide_name} not found in LMDB metadata.")
    num_patches = pn_dict[slide_name]
    
    # Generate patch IDs
    patch_ids = [f"{slide_name}-{i}" for i in range(num_patches)]

    # Allocate memory for patches (adjust dimensions and dtype as needed)
    # Assuming patches are 224x224, 3 channels, and will be normalized later
    patches_tensor = torch.empty((len(patch_ids), 3, 224, 224), dtype=torch.float32)

    # Load and decode data into torch.tensor
    for i, key_str in enumerate(patch_ids):
        patch_bytes = txn.get(key_str.encode('ascii'))
        if patch_bytes is None:
            print(f"Warning: Key {key_str} not found in LMDB.")
            continue
        # Assuming the stored value is pickled image bytes
        img_array = imfrombytes(pickle.loads(patch_bytes).tobytes()) # Or .tobytes() if it's already bytes
        patches_tensor[i] = torch.from_numpy(img_array.transpose(2, 0, 1)) # HWC to CHW

# Normalize the data (example using ImageNet stats)
# Ensure values are in [0, 255] before this normalization if they aren't already
mean = torch.tensor([0.485, 0.456, 0.406]).view((1, 3, 1, 1)) * 255.0
std = torch.tensor([0.229, 0.224, 0.225]).view((1, 3, 1, 1)) * 255.0

# If your patches_tensor is already in [0,1] range, remove * 255.0 from mean/std
# If your patches_tensor is uint8 [0,255], convert to float first: patches_tensor.float()
patches_tensor = (patches_tensor.float() - mean) / std

env.close()
```

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**
```
@misc{tang2025revisitingendtoendlearningslidelevel,
      title={Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology}, 
      author={Wenhao Tang and Rong Qin and Heng Fang and Fengtao Zhou and Hao Chen and Xiang Li and Ming-Ming Cheng},
      year={2025},
      eprint={2506.02408},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.02408}, 
}
```