README.md

metadata

license: cc-by-4.0

SMMILe: WSI Embedding Superpixel Data, and Checkpoints

This repository contains embedding superpixel segmentation and checkpoint data for six publicly available whole-slide image (WSI) datasets. The embeddings were generated using two different encoders: ResNet-50 and Conch. Each dataset is available as a ZIP file containing both the embeddings and the corresponding superpixel segmentation results.

Datasets and Files

Checkpoints of SMMILe for different datasets and encoders are all saved in ckpt_submit.zip.

The rest of the ZIP files contain the embeddings and superpixel segmentations for one dataset, which are processed using either ResNet-50 or Conch. The files are stored using Git LFS due to their large size.

File Overview:

Dataset	Encoder	File Name	Size
Breast	Conch	camel_conch.zip	5.76 GB
Breast	ResNet-50	camel_res50.zip	17.5 GB
Gastric	Conch	gastric_conch.zip	489 MB
Gastric	ResNet-50	gastric_res50.zip	1.13 GB
Lung	Conch	lung_conch.zip	1.4 GB
Lung	ResNet-50	lung_res50.zip	3.12 GB
Ovarian	Conch	ovarian_conch.zip	2.5 GB
Ovarian	ResNet-50	ovarian_res50.zip	6.42 GB
Prostate	Conch	prostate_conch.zip	35.6 MB
Prostate	ResNet-50	prostate_res50.zip	69.1 MB
Renal-3	Conch	renal3_conch.zip	1.47 GB
Renal-3	ResNet-50	renal3_res50.zip	3.2 GB

Structure of Each ZIP File:

Each ZIP file contains two folders:

• Model Embeddings: Embeddings for patches extracted from WSIs, stored as .npy files.
• Superpixel Segmentations: Corresponding superpixel segmentation results for each WSI, stored as .npy files.

Example File Structure:

camel_res50.zip
  ├── resnet50/            # ResNet-50 embeddings (1024-dimensional)
  └── sp_xxx/              # Superpixel segmentations

gastric_conch.zip
  ├── conch/               # Conch embeddings (512-dimensional)
  └── sp_xxx/              # Superpixel segmentations

Embeddings (.npy Files):

Keys in the Embedding Files:

• 'feature': Embedding vectors of image patches.
• 'inst_label': Annotated label of each patch. A value of -1 indicates a non-annotated patch.
• 'index': Coordinates and size of the patch in the format x_y_size, where x and y represent the top-left coordinates.

Superpixel Segmentations (.npy Files):

Keys in the Superpixel Files:

• 'm_slic': Superpixel index for each patch.
• 'm_adj': Adjacency matrix representing the relationships between superpixels.