README.md

---
license: cc-by-nc-nd-4.0
tags:
- medical-imaging
- chest-xray
- embeddings
- mimic-cxr
- coreset-selection
- quantum-machine-learning
---

# MIMIC-CXR Embeddings Dataset

This dataset contains pre-extracted embeddings from MIMIC-CXR chest X-ray images using multiple state-of-the-art vision models. The embeddings are organized by coreset selection strategies for efficient training of quantum machine learning models.

## Dataset Overview

- **Source**: MIMIC-CXR Database
- **Total Seeds**: 20 (seed_0 through seed_19)
- **Coreset Strategies**: 3 per seed
- **Embedding Models**: 5 vision transformer architectures
- **Total Samples**: ~1,999–2,372 samples per strategy (varies by seed)
- **File Format**: Parquet (and legacy Pickle for CLIP-BioMed)

## Coreset Selection Strategies

Each seed contains three coreset selection strategies. Sample counts shown are for seed_0; exact counts vary slightly by seed:

| Strategy | Name | Samples (seed_0) | Description |
|----------|------|-------------------|-------------|
| **5** | PathologyStratifiedClean | 1,999 | Stratified sampling based on pathology labels |
| **9** | GradMatch | 2,371 | Gradient matching for representative subset selection |
| **11** | Uncertainty | 2,371 | Uncertainty-based active learning sample selection |

## Embedding Types (ViT-16 and ViT-32)

For ViT-Base-Patch16-224 and ViT-Base-Patch32-224, two embedding variants are provided per data type, distinguished by filename suffix:

### `_cls_embedding` — CLS Token Embedding
The standard 768-dim representation extracted from the `[CLS]` token of the final transformer layer. This is the model's global summary vector used in classification tasks.

### `_gap_embedding` — Multi-Layer Global Average Pooling
A richer 768-dim representation computed by pooling patch token hidden states across the last 4 transformer layers:

1. Extract patch token hidden states from the **last 4 transformer blocks** (CLS token excluded)
2. Stack into shape `[4, num_patches, 768]`
3. **Mean-pool across the layer dimension** → `[num_patches, 768]`
4. **Mean-pool across the patch dimension** → `[768]`

| Model | Patch tokens per image | Layers pooled |
|-------|----------------------|---------------|
| ViT-Base-Patch16-224 | 196 (14 × 14) | Last 4 of 12 |
| ViT-Base-Patch32-224 | 49 (7 × 7) | Last 4 of 12 |

## Embedding Models

### 1. CLIP-BioMed
- **Path**: `clip-biomed-embeddings/`
- **Format**: Pickle (`.pkl`)
- **Files**:
  - `data_type5_insurance.pkl` (1,999 samples)
  - `data_type9_insurance_2371rows.pkl` (2,371 samples)
  - `data_type11_insurance_2371rows.pkl` (2,371 samples)

### 2. MedSigLIP-448
- **Path**: `medsiglip-448-embeddings/`
- **Format**: Parquet (`.parquet`)
- **Embedding**: CLS token (1,152-dim via `google/siglip-so400m-patch14-384`)
- **Files** (seed_0):
  - `data_type5_n1999_seed0_medsiglip_448.parquet` (1,999 samples)
  - `data_type9_n2371_seed0_medsiglip_448.parquet` (2,371 samples)
  - `data_type11_n2371_seed0_medsiglip_448.parquet` (2,371 samples)

### 3. ViT-Base-Patch32-224
- **Path**: `vit-base-patch32-224-embeddings/`
- **Format**: Parquet (`.parquet`)
- **Embedding dimension**: 768
- **Files** (seed_0):
  - `data_type5_n1999_seed0_vit_base_patch32_224_cls_embedding.parquet` — CLS token
  - `data_type5_n1999_seed0_vit_base_patch32_224_gap_embedding.parquet` — Multi-layer GAP
  - `data_type9_n2371_seed0_vit_base_patch32_224_cls_embedding.parquet`
  - `data_type9_n2371_seed0_vit_base_patch32_224_gap_embedding.parquet`
  - `data_type11_n2371_seed0_vit_base_patch32_224_cls_embedding.parquet`
  - `data_type11_n2371_seed0_vit_base_patch32_224_gap_embedding.parquet`

### 4. ViT-Base-Patch16-224
- **Path**: `vit-base-patch16-224-embeddings/`
- **Format**: Parquet (`.parquet`)
- **Embedding dimension**: 768
- **Files** (seed_0):
  - `data_type5_n1999_seed0_vit_base_patch16_224_cls_embedding.parquet` — CLS token
  - `data_type5_n1999_seed0_vit_base_patch16_224_gap_embedding.parquet` — Multi-layer GAP
  - `data_type9_n2371_seed0_vit_base_patch16_224_cls_embedding.parquet`
  - `data_type9_n2371_seed0_vit_base_patch16_224_gap_embedding.parquet`
  - `data_type11_n2371_seed0_vit_base_patch16_224_cls_embedding.parquet`
  - `data_type11_n2371_seed0_vit_base_patch16_224_gap_embedding.parquet`

### 5. RAD-DINO
- **Path**: `rad-dino-embeddings/20-seeds/seed_<N>/`
- **Format**: Parquet (`.parquet`)
- **Files** (seed_0):
  - `data_type5_n1998_seed0_rad_dino.parquet` (1,998 samples)
  - `data_type9_n2370_seed0_rad_dino.parquet` (2,370 samples)
  - `data_type11_n2370_seed0_rad_dino.parquet` (2,370 samples)
- **Note**: 1 sample missing per strategy (see Verification Status below)

## Folder Structure

```
qml-mimic-cxr-embeddings/
├── coreset-ids/
│   ├── seed_0/
│   │   ├── coreset-has_pathology-5-PathologyStratifiedClean-seed_0.txt
│   │   ├── coreset-has_pathology-9-GradMatch-seed_0.txt
│   │   └── coreset-has_pathology-11-Uncertainty-seed_0.txt
│   └── seed_1/ ... seed_19/
├── clip-biomed-embeddings/
│   ├── README.md
│   ├── data_type5_insurance.pkl
│   ├── data_type9_insurance_2371rows.pkl
│   ├── data_type11_insurance_2371rows.pkl
│   ├── data-cleaned-pca-100/
│   │   ├── data_type5_insurance.pkl
│   │   ├── data_type9_insurance.pkl
│   │   ├── data_type9_insurance_2371rows.pkl
│   │   ├── data_type11_insurance.pkl
│   │   ├── data_type11_insurance_2371rows.pkl
│   │   └── models/
│   │       ├── global_stats_100_type{9,11}_2371rows.npz
│   │       └── svd_components_100_type{9,11}_2371rows.npz
│   ├── data-cleaned-pca-500/
│   │   └── (same structure as pca-100, with 500-dim variants)
│   ├── data-cleaned-pca-1000/
│   │   └── (same structure, with 1000-dim variants)
│   ├── data-cleaned-pca-1999/
│   │   └── (same structure, with 1999-dim variants)
│   └── 20-seeds/
│       ├── seed_0/
│       │   ├── data_type5_n1999.parquet
│       │   ├── data_type9_n2371.parquet
│       │   └── data_type11_n2371.parquet
│       └── seed_1/ ... seed_19/
├── medsiglip-448-embeddings/
│   ├── data_type5_n1999_seed0_medsiglip_448.parquet
│   ├── data_type9_n2371_seed0_medsiglip_448.parquet
│   ├── data_type11_n2371_seed0_medsiglip_448.parquet
│   └── 20-seeds/
│       ├── seed_0/
│       │   ├── data_type5_n1999.parquet
│       │   ├── data_type9_n2371.parquet
│       │   └── data_type11_n2371.parquet
│       └── seed_1/ ... seed_19/
├── vit-base-patch32-224-embeddings/
│   ├── data_type5_n1999_seed0_vit_base_patch32_224_cls_embedding.parquet
│   ├── data_type5_n1999_seed0_vit_base_patch32_224_gap_embedding.parquet
│   ├── data_type9_n2371_seed0_vit_base_patch32_224_cls_embedding.parquet
│   ├── data_type9_n2371_seed0_vit_base_patch32_224_gap_embedding.parquet
│   ├── data_type11_n2371_seed0_vit_base_patch32_224_cls_embedding.parquet
│   ├── data_type11_n2371_seed0_vit_base_patch32_224_gap_embedding.parquet
│   └── 20-seeds/
│       ├── seed_0/
│       │   ├── data_type5_n1999_cls_embedding.parquet
│       │   ├── data_type5_n1999_gap_embedding.parquet
│       │   ├── data_type9_n2371_cls_embedding.parquet
│       │   ├── data_type9_n2371_gap_embedding.parquet
│       │   ├── data_type11_n2371_cls_embedding.parquet
│       │   └── data_type11_n2371_gap_embedding.parquet
│       └── seed_1/ ... seed_19/
├── vit-base-patch16-224-embeddings/
│   ├── data_type5_n1999_seed0_vit_base_patch16_224_cls_embedding.parquet
│   ├── data_type5_n1999_seed0_vit_base_patch16_224_gap_embedding.parquet
│   ├── data_type9_n2371_seed0_vit_base_patch16_224_cls_embedding.parquet
│   ├── data_type9_n2371_seed0_vit_base_patch16_224_gap_embedding.parquet
│   ├── data_type11_n2371_seed0_vit_base_patch16_224_cls_embedding.parquet
│   ├── data_type11_n2371_seed0_vit_base_patch16_224_gap_embedding.parquet
│   └── 20-seeds/
│       ├── seed_0/
│       │   ├── data_type5_n1999_cls_embedding.parquet
│       │   ├── data_type5_n1999_gap_embedding.parquet
│       │   ├── data_type9_n2371_cls_embedding.parquet
│       │   ├── data_type9_n2371_gap_embedding.parquet
│       │   ├── data_type11_n2371_cls_embedding.parquet
│       │   └── data_type11_n2371_gap_embedding.parquet
│       └── seed_1/ ... seed_19/
├── rad-dino-embeddings/
│   └── 20-seeds/
│       ├── seed_0/
│       │   ├── data_type5_n1998_seed0_rad_dino.parquet
│       │   ├── data_type9_n2370_seed0_rad_dino.parquet
│       │   └── data_type11_n2370_seed0_rad_dino.parquet
│       └── seed_1/ ... seed_19/
└── tests/
    ├── README.md
    ├── verify_all_embeddings.py
    ├── verify_basic_embeddings.py
    └── verify_rad_dino.py
```

## Data Format

Parquet files (ViT-16, ViT-32, MedSigLIP, RAD-DINO) contain a pandas DataFrame where:
- **`embedding`**: Pre-extracted feature vector (as a list of floats) from the respective model/variant
- **Metadata columns**: `dicom_id`, `subject_id`, `study_id`, and additional MIMIC-CXR metadata

Pickle files (CLIP-BioMed) follow the same structure.

### Loading Example

```python
import pandas as pd

# Load a CLS embedding (parquet)
df = pd.read_parquet(
    'vit-base-patch16-224-embeddings/'
    'data_type5_n1999_seed0_vit_base_patch16_224_cls_embedding.parquet'
)
embeddings = df['embedding'].tolist()  # list of 768-dim vectors

# Load a GAP embedding (parquet)
df_gap = pd.read_parquet(
    'vit-base-patch16-224-embeddings/'
    'data_type5_n1999_seed0_vit_base_patch16_224_gap_embedding.parquet'
)
gap_embeddings = df_gap['embedding'].tolist()  # list of 768-dim vectors

# Load from HuggingFace Hub directly
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id='MITCriticalData/qml-mimic-cxr-embeddings',
    filename='vit-base-patch16-224-embeddings/data_type9_n2371_seed0_vit_base_patch16_224_gap_embedding.parquet',
    repo_type='dataset'
)
df = pd.read_parquet(path)
```

## Verification Status (seed_0)

All seed_0 coreset IDs have been verified against extracted embeddings:

| Embedding Type | Strategy 5 | Strategy 9 | Strategy 11 |
|----------------|------------|------------|-------------|
| **CLIP-BioMed** | ✓ 100% (1,999/1,999) | ✓ 100% (2,371/2,371) | ✓ 100% (2,371/2,371) |
| **MedSigLIP-448** | ✓ 100% (1,999/1,999) | ✓ 100% (2,371/2,371) | ✓ 100% (2,371/2,371) |
| **ViT-Patch32 CLS** | ✓ 100% (1,999/1,999) | ✓ 100% (2,371/2,371) | ✓ 100% (2,371/2,371) |
| **ViT-Patch32 GAP** | ✓ 100% (1,999/1,999) | ✓ 100% (2,371/2,371) | ✓ 100% (2,371/2,371) |
| **ViT-Patch16 CLS** | ✓ 100% (1,999/1,999) | ✓ 100% (2,371/2,371) | ✓ 100% (2,371/2,371) |
| **ViT-Patch16 GAP** | ✓ 100% (1,999/1,999) | ✓ 100% (2,371/2,371) | ✓ 100% (2,371/2,371) |
| **RAD-DINO** | ✓ 99.95% (1,998/1,999) | ✓ 99.96% (2,370/2,371) | ✓ 99.96% (2,370/2,371) |

**RAD-DINO Missing Samples:**
- Strategy 5: `db806824-34de7587-691208b6-19301aaa-15cca66c`
- Strategy 9: `1d413540-516c7ce1-0a64dfe2-78c7b93e-808b2fce`
- Strategy 11: `669089f6-b0ff4487-f652652d-80e2925d-7e2b2511`

## License

This dataset is released under **CC-BY-NC-ND-4.0** (Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International).