MediaEval Medico 2026 - Task 1

Team Info

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

Model Overview

This submission uses CATA-Qwen3B, a curriculum-aware 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. The model was trained with curriculum alignment warmup and gate regularization before full-data continuation.

Architecture Summary

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
→ curriculum-gated TopoAdapter injection in the last 8 decoder layers
→ generated answer

CATA-Qwen3B extends the full structural TopoAdapter architecture with a curriculum-aware training recipe. The first stage warms up language adaptation, while later training restores spatial/topological alignment losses and gate regularization. At inference time, the extracted structural features condition both visual fusion and the Qwen decoder TopoAdapters.

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

submission_task1.py     Main inference script
requirements.txt        Python dependencies
checkpoints/last.pt     Trainable checkpoint weights
src/                    Model, topology, and post-processing code