7 open VLMs were tested as a local second reader for medical de-identification
MaziyarPanahi · x · 2026-07-22
We tested 7 small open VLMs as a local second-reader for de-identifying medical images.
- Qwen3-VL 4B: 7/7 identifiers, 6.9s, 0 false alarms — chosen as the seat.
- Gemma 4 E2B: 6/7.
- Qwen3-VL 2B: 4/7, but the fastest.
- NVIDIA LocateAnything-3B: perfect bounding boxes, but it never judged what was personal.
- InternVL3 2B: correctly read the patient’s name, then still said it was not personal.
The takeaway: reading is easy; deciding is the hard part. The follow-up example showed OpenMed’s PII model passing a chest X-ray, while Qwen3-VL caught a clinic name still burned into the image. Both models are open on Hugging Face. The question posed: should every scan get a second reader?
Related event: Qwen3-VL Beats 6 Open-Source VLMs in Medical Image Desensitization Test(3 posts)→
More from Multimodal
- NVIDIA’s Cosmos 3 Super claims 25x faster image and video generation — NVIDIAAI · 2026-07-22
- A creator updates a city-pop LoKr adapter for Krea with new samples — lacerating_aura · 2026-07-22
- AI-Generated Game Worlds May Be Impressive for 30 Seconds, Not an Hour — Slight_Control9311 · 2026-07-22
- A browser black-hole demo built with Gemini 3.6 Flash runs in three files at 60 FPS — prasenx · 2026-07-22
- RunwayML Agent generates a piece titled “The Bundyssey” — tlakomy · 2026-07-22
- Storyboard Reference Studio turns clips into storyboards in seconds — bennash · 2026-07-22