You can deploy this model and start earning money today!

minhquang47/
medico2026-task1-topofusion-qwen3b

Name: medico2026-task1-topofusion-qwen3b
Availability: InStock

Views

No views yet

MediaEval Medico 2026 - Task 1

Team Info

Team name: Sweet&Sour
Member: Nguyễn Minh Quang
Email: nmquang04072005@gmail.com

Model Overview

This submission uses TopoFusion-Qwen3B, a structural VQA system that combines a frozen ViT visual encoder, online lesion-prior/topological feature extraction, full structural fusion, and Qwen2.5-3B-Instruct with QLoRA r16 plus deep TopoAdapter layers to generate textual answers.

Architecture Summary

text

Image
→ ViT visual encoder
→ online structural extraction:
   - lesion prior mask M_prior
   - topological feature map
   - global morphology features
→ structural visual fusion
→ Qwen2.5-3B-Instruct + QLoRA r16
→ deep TopoAdapter injection in the last 8 decoder layers
→ generated answer

TopoFusion-Qwen3B uses the full structural signal directly during inference. The lesion prior and topological descriptors are used both by the visual fusion module and by the deep TopoAdapter condition vector inside the Qwen decoder.

Dependencies

Install the required packages with:

pip install -r requirements.txt

The script also loads the base LLM from Hugging Face:

Qwen/Qwen2.5-3B-Instruct

How to Run

From this folder, run:

python submission_task1.py

The system will load the checkpoint from:

checkpoints/last.pt

It will run prediction on the Task 1 test split and write the output file:

predictions_1.json

Files