Скачать или смотреть Medico 2025: Curriculum-Guided Fine-Tuning for Multimodal VQA in GI Endoscopy (Team Lama4Vision)

Curriculum-Guided Fine-Tuning for Multimodal VQA in GI Endoscopy (Team Lama4Vision)

Authors: Mahdi Azmoodeh-Kalati, Mohammad Sadegh Maghareh, Saba Alavi and Reza Lashgari

Presenter: Mahdi Azmoodeh-Kalati

Event: MediaEval Medico 2025 Challenge
Task: Visual Question Answering (with multimodal explanations) for Gastrointestinal Imaging

This video presents the team’s contribution to the MediaEval Medico 2025 Challenge, a benchmark focused on medical Visual Question Answering (VQA) using gastrointestinal (GI) endoscopy data. The work is evaluated using the Kvasir-VQA-x1 dataset, which includes 159k+ multimodal question–answer pairs designed to support clinical reasoning, explainability, and robustness.

The participants describe their system design, methodology, and experimental results for Subtask 1 (AI performance on GI-VQA) and/or Subtask 2 (clinician-oriented multimodal explanations). Approaches typically include vision–language model architectures, image–text fusion strategies, medical reasoning techniques, and methods for generating explainable predictions such as visual heatmaps, segmentation masks, or structured textual rationales.

This presentation was recorded for the MediaEval Workshop, held on 25–26 October 2025 in Dublin, Ireland and Online, as part of the MediaEval 2025 Benchmarking Initiative.

For full challenge details, papers, datasets, scripts, and leaderboard entries, please visit:
GitHub Repository: https://github.com/simula/MediaEval-M...

Competition Results: https://github.com/simula/MediaEval-M...

Recordings Playlist:    • MediaEval Medico 2025  

Thank you to all contributors, teams, and reviewers for advancing research in explainable and trustworthy medical AI.