Benchmark & Challenges

Benchmark & Challenges

benchmarkchallengedatasetretinal

Overview

Progress in medical AI is only reproducible when the community shares standardised datasets, evaluation protocols, and leaderboards. I contribute to this infrastructure by co-organising international challenges and building open benchmarks that allow fair comparison of methods under identical conditions.

MARIO Challenge — MICCAI 2024

The MARIO (Monitoring Age-Related Macular Degeneration Intelligence and Outcomes) challenge was a satellite event at MICCAI 2024 in Marrakesh. It provided:

  • A multi-modal longitudinal OCT dataset for AMD progression prediction
  • Standardised train/validation/test splits with held-out labels for fair evaluation
  • Two tracks: binary conversion prediction and interval-to-conversion regression
  • Participation from international teams across 3 continents

Results were presented at the MICCAI 2024 main conference. The challenge data and evaluation server remain publicly available for continued benchmarking.

Challenge outcomes

The challenge revealed several key insights:

  1. Multi-modal fusion consistently outperformed single-modality approaches
  2. Longitudinal models with ≥12 months of history yielded the largest gains
  3. Uncertainty calibration was a significant differentiator between top-performing systems

Why challenges matter

  • Reproducibility: all participants evaluate on the same held-out test set
  • Community progress: leaderboards incentivise focused innovation
  • Clinical translation: challenge metrics are designed in collaboration with clinicians to ensure clinical relevance
  • Data sharing: challenges make rare clinical datasets accessible under controlled conditions
← Back to Research