Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- chest-xray
|
| 4 |
+
- radiology
|
| 5 |
+
- report-generation
|
| 6 |
+
- mimic-cxr
|
| 7 |
+
license: apache-2.0
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# LAPVQA — Radiology Report Generation (Frozen Off-the-shelf Encoders)
|
| 11 |
+
|
| 12 |
+
Part of the [LAPVQA collection](https://huggingface.co/collections/dmusingu/lapvqa).
|
| 13 |
+
|
| 14 |
+
## Description
|
| 15 |
+
|
| 16 |
+
Generative decoder heads for **Radiology Report Generation (RRG)** on MIMIC-CXR,
|
| 17 |
+
trained on top of five **frozen** off-the-shelf vision encoders.
|
| 18 |
+
Each file contains the decoder weights; the encoder is not bundled.
|
| 19 |
+
|
| 20 |
+
## Results (MIMIC-CXR test set)
|
| 21 |
+
|
| 22 |
+
| Encoder | BLEU-4 | ROUGE-L | RadGraph-s |
|
| 23 |
+
|---|---|---|---|
|
| 24 |
+
| CLIP ViT-L/14 | 0.034 | 0.168 | 0.197 |
|
| 25 |
+
| Florence-2 | 0.035 | 0.169 | 0.205 |
|
| 26 |
+
| SigLIP | 0.036 | 0.168 | 0.211 |
|
| 27 |
+
| CoCa | 0.030 | 0.160 | 0.193 |
|
| 28 |
+
| OWLv2 | 0.034 | 0.169 | 0.197 |
|
| 29 |
+
|
| 30 |
+
SigLIP achieves the best RadGraph-s (0.211) among all frozen encoders, consistent with
|
| 31 |
+
its larger patch-token count (729 tokens at 384 px) providing finer spatial detail.
|
| 32 |
+
|
| 33 |
+
## Files
|
| 34 |
+
|
| 35 |
+
| File | Encoder backbone |
|
| 36 |
+
|---|---|
|
| 37 |
+
| `clip-vit-l14.pt` | CLIP ViT-L/14 |
|
| 38 |
+
| `siglip.pt` | SigLIP |
|
| 39 |
+
| `florence2.pt` | Florence-2 |
|
| 40 |
+
| `coca.pt` | CoCa |
|
| 41 |
+
| `owlv2.pt` | OWLv2 |
|