PancCADx is an interpretable multimodal framework for pancreatic cancer diagnosis via endoscopic ultrasound (EUS). This adapter was trained using a two-stage alignment strategy:

SFT (Supervised Fine-Tuning): Learning diagnostic patterns from expert annotations
DPO (Direct Preference Optimization): Error-driven alignment to reduce hallucinations

Usage

python

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from peft import PeftModel

# Load base model
base_model = Qwen3VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-8B-Thinking",
    torch_dtype="auto",
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "shan1984/PancCADx-DPO")
processor = AutoProcessor.from_pretrained("shan1984/PancCADx-DPO")

Training Details

Base model: Qwen3-VL-8B-Thinking
LoRA config: rank=128, alpha=256, target=all
DPO: beta=0.3, lr=1e-7, epochs=10
Training framework: LLaMA-Factory

Performance (External Validation, n=191)

Metric	Value
Sensitivity	95.74%
Specificity	77.78%
Accuracy	89.53%

Citation

bibtex

@inproceedings{hu2026panccadx,
  title={PancCADx: A Multimodal Framework for Pancreatic Cancer Diagnosis},
  author={Hu, Shan and Xiao, Changhong and Qin, Xianzheng and Mei, Bin and Cheng, Bin and Wang, Zhongyuan},
  booktitle={MICCAI},
  year={2026}
}

License

Apache 2.0