SCIN-Dermatology-Generational-Embeddings

This dataset contains pre-extracted image-level feature embeddings and clinical/demographic metadata for 6,517 photographs (grouped into 3,061 cases). These features were extracted using two distinct encoder backbones:

1. TIPSv2 (Gen 2): 1,024-dimensional feature embeddings extracted from the open-weights google/tipsv2-l14 visual backbone.
2. Gemma4 (Gen 3): 2,560-dimensional visual token embeddings extracted from the on-device litert-community/gemma-4-E4B-it-litert-lm multimodal visual encoder.

Storing pre-extracted embeddings enables fast training of downstream classifiers—such as Multiple Instance Learning (MIL) models—without requiring local vision model forward passes.

Dataset Structure

The dataset contains two compressed Apache Parquet files:

.
├── README.md
├── tipsv2_embeddings.parquet    # Columnar 1024-d embeddings + metadata
└── gemma4_embeddings.parquet    # Columnar 2560-d embeddings + metadata

Schema Structure

Each Parquet file follows the exact same schema:

Column Name	Data Type	Description
case_id	string	Unique patient case ID
image_id	string	Unique image ID
embedding	list[float]	Numeric feature vector (1,024-d for TIPSv2, 2,560-d for Gemma4)
dataset_split	string	Stratified split assignment (train, val, test)
binary_label	int	Triage label (0 = non-infectious, 1 = infectious)
primary_diagnosis	string	Primary consensus dermatology diagnosis (e.g. Eczema)
fitzpatrick_skin_type	string	Fitzpatrick skin type grading (FST1–FST6)
monk_skin_tone_us	float	Monk Skin Tone grading (US standard, scale 1–10)
monk_skin_tone_india	float	Monk Skin Tone grading (India standard, scale 1–10)

Quick Start (Python)

To load the embeddings and meta columns directly using the Hugging Face datasets library:

from datasets import load_dataset

# Load TIPSv2 embeddings
tipsv2_dataset = load_dataset(
    "HawkFranklin-Research/SCIN-Dermatology-Generational-Embeddings", 
    data_files="tipsv2_embeddings.parquet"
)

# Load Gemma4 embeddings
gemma4_dataset = load_dataset(
    "HawkFranklin-Research/SCIN-Dermatology-Generational-Embeddings", 
    data_files="gemma4_embeddings.parquet"
)

# Access a sample vector
sample = tipsv2_dataset["train"][0]
print(sample["case_id"])  # e.g., '-1000600354148496558'
print(sample["primary_diagnosis"])  # e.g., 'Eczema'
print(len(sample["embedding"]))  # 1024

Citation & License

This dataset is distributed under the MIT License. The underlying image classifications and clinical features are derived from the Google SCIN repository. If you use this dataset in your research, please cite the primary SCIN work:

@article{Ward2024_SCIN_JAMANetworkOpen,
  author    = {Ward, Edwin and others},
  title     = {Skin Condition Image Network (SCIN): A diverse dataset of patient-submitted skin photographs},
  journal   = {JAMA Network Open},
  year      = {2024},
  volume    = {7},
  number    = {5},
  pages     = {e2410389}
}