from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration
from PIL import Image

processor = PaliGemmaProcessor.from_pretrained("MedeHealth/paligemma-chest-xray-vision")
model = PaliGemmaForConditionalGeneration.from_pretrained("MedeHealth/paligemma-chest-xray-vision")

image = Image.open("chest_xray.png")
prompt = "Analyze this chest X-ray and provide a structured radiology report."

inputs = processor(text=prompt, images=image, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=1024)
report = processor.decode(outputs[0], skip_special_tokens=True)
print(report)

Deployment

Base Model: google/medgemma-4b-it
Task: Medical image report generation (image-text-to-text)
Language: Portuguese (Brazilian)
Fine-tuning Method: QLoRA (4-bit quantization)

Deploy to HuggingFace Inference Endpoints with task type: image-text-to-text

Limitations

For research and educational purposes
Should not be used as sole basis for clinical decisions
Requires radiologist review and validation