← Demos
MEDSCAN
// encrypted medical image retrieval
CKKS
128-BIT
CC MODE
Search radiology reports without exposing the patient image.
Upload a chest X-ray or describe findings in text. The query is encrypted with Fully Homomorphic Encryption (CKKS) and matched against 3,000+ encrypted radiology reports from the IU X-Ray dataset. The server never sees the query or the database in plaintext.
1 Upload X-ray or type a clinical query
2 BiomedCLIP encodes the query into a 512-dim embedding
3 Embedding is encrypted (CKKS) and searched against encrypted DB
4 Matching radiology reports are returned with similarity scores
Hospital A
<-->
Medical Center B
query input
Drop chest X-ray here or click to upload
SAMPLE X-RAYS
encryption pipeline
Encrypting query...
Upload an X-ray image or enter a text query to search.
datasheet
| Dataset | |
|---|---|
| Name | IU X-Ray (Indiana University Chest X-rays) |
| Source | HuggingFace ykumards/open-i |
| Size | 3,666 image-report pairs |
| Modalities | Frontal chest X-ray (JPEG) + radiology findings (free text) |
| Annotations | MeSH terms, clinical problems, indication, impression |
| Embedding Model | |
| Name | BiomedCLIP-PubMedBERT_256-vit_base_patch16_224 |
| Publisher | Microsoft Research |
| Architecture | PubMedBERT (text) + ViT-B/16 (vision) |
| Dimension | 512 |
| Similarity | Cosine (L2-normalized embeddings) |
| Encryption | |
| Scheme | CKKS |
| Security | 128-bit (quantum-safe lattice) |
| Mode | CC — both query and database vectors encrypted |
| Library | libevi (CryptoLab HEaaN) |
| Compute Environment | |
| Backend | CPU only (HEM CPU backend) |
| Threads | 8 (OMP_NUM_THREADS=8) |
| Embedding | BiomedCLIP on CPU (~1s per query) |
| Encrypted Search | ~300ms for 3,666 encrypted vectors |