Radmatch-M3-V1
ranking radiology impressions
2 min readupdated
Designed for English medical texts, this model ranks matching diagnostic impressions based on a study's findings section. Built on BAI/bge-m3 and fine-tuned for semantic search in radiology, it doesn't generate new reports—instead, it surfaces relevant phrasing from existing options. It's a solid fit for researchers and developers working with medical text data.
Key features
- Tailored for radiology texts (abdominal, chest, neurology, and more)
- 1024-dimensional dense embeddings for semantic search
- Higher accuracy on internal benchmarks compared to base models
- Full fine-tuning of the BAI/bge-m3 encoder
What you can do
- Set up search across a radiology impression database in a research system
- Rank candidate impressions based on imaging findings
- Test dense retrieval using the FlagEmbedding library
Example
Sample code for encoding study findings and ranking candidates:
import numpy as np
from FlagEmbedding import BGEM3FlagModel
model = BGEM3FlagModel(
"RadiOtter/Radmatch-M3-V1",
use_fp16=False,
)
findings = ["The lungs are clear. No pleural effusion or pneumothorax."]
impressions = [
"No acute cardiopulmonary abnormality.",
"Left lower lobe pneumonia with a small pleural effusion.",
"Right-sided pneumothorax."
]
queries = model.encode(findings, batch_size=1, max_length=512)["dense_vecs"]
passages = model.encode(impressions, batch_size=3, max_length=256)["dense_vecs"]
queries = np.asarray(queries, dtype=np.float32)
passages = np.asarray(passages, dtype=np.float32)
queries /= np.maximum(np.linalg.norm(queries, axis=1, keepdims=True), 1e-12)
passages /= np.maximum(np.linalg.norm(passages, axis=1, keepdims=True), 1e-12)
scores = queries @ passages.T
for index in np.argsort(-scores[0]):
print(f"{scores[0, index]:.4f} {impressions[index]}")Read more
DetailsBuilt on BAI/bge-m3 (0.57B parameters).
Architecture: XLM-RoBERTa-based BGE-M3 encoder.
Limitations: not a diagnostic tool and has not undergone independent clinical validation.










Comments