Qwen3.8-27B-Radiology-Impression

🩻 Qwen 3.8-27B Clinical Radiology Impression & Consultation AI

<div align="center">

</div>

---

🌟 Overview & Key Links

Medico/Qwen3.8-27B-Radiology-Impression is a dense 27-Billion parameter clinical foundation model fine-tuned on 659,381 real-world clinical findings-to-impression pairs across diverse imaging modalities (CT, MRI, Chest Radiography, Ultrasound, and PET/CT).

• 🌐 Live Interactive Web Demo: Hugging Face Space Demo
• 📊 Curated Training Corpus: Medico/radiology-reports-curated
• ⚡ Base Architecture: Qwen/Qwen3.8-27B (Dense 27B model, qwen3_5 architecture)

---

🚀 Free Unlimited Chat & Deployment Options

You can chat with this model with unlimited access for free using any of the following methods:

Option 1: Live Web App (Instant Free Access)

Open the official Hugging Face Space: 👉 https://huggingface.co/spaces/Medico/Qwen3.8-27B-Radiology-Impression-Demo

Option 2: Free 1-Line Local Chat (Ollama)

Run directly on your local machine with GPU/CPU:

ollama run hf.co/Medico/Qwen3.8-27B-Radiology-Impression

Option 3: Free Cloud GPU Inference (Google Colab / Kaggle / Lightning AI)

Run this Python snippet on any free GPU environment (Tesla T4, L4, A10G):

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen3.8-27B"
LORA_REPO = "Medico/Qwen3.8-27B-Radiology-Impression"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True
)

tokenizer = AutoTokenizer.from_pretrained(LORA_REPO, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
model = PeftModel.from_pretrained(base_model, LORA_REPO)

# Clinical Prompting
findings = """
CTA CHEST: Large saddle pulmonary embolus straddling the bifurcation with RV/LV ratio of 1.4 and interventricular flattening.
"""

messages = [
    {"role": "system", "content": "You are an expert diagnostic radiologist. Generate a structured clinical IMPRESSION."},
    {"role": "user", "content": f"FINDINGS:\n{findings}"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.2)

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

---

🎯 Clinical Benchmark Capabilities

1. Findings-to-Impression Synthesis: Condenses verbose imaging observations into clear, prioritized clinical impressions.
2. Critical Findings & Urgent Alerts: Highlights acute conditions (pulmonary embolism, large vessel occlusion, aortic dissection, tension pneumothorax, appendicitis).
3. Structured Staging & Reporting: Aligned with ACR Fleischner Society criteria, BI-RADS, LI-RADS, and PI-RADS reporting frameworks.

---

🔒 Clinical Disclaimer

This model is released for research, clinical decision support benchmarking, and medical NLP exploration. It should not replace independent radiological interpretation by certified medical professionals.