Medical imaging AI trained predominantly on Western datasets performs poorly on Asian patient populations. Bone density distributions differ, disease prevalence differs, and imaging protocols differ across institutions. This is not a hypothesis — we've measured it.
In our scoliosis Cobb angle measurement project (published in Annals of Academy of Medicine Singapore, 2024), models trained on international benchmarks underperformed by 12% on local patient radiographs. After fine-tuning on local data, performance matched or exceeded published benchmarks.
Practical approaches to Asian medical imaging AI:
- Local data is irreplaceable — A 1,000-case local dataset outperforms transfer learning from 100,000 Western cases for Singapore population-specific conditions.
- Multi-institution federation — We built a federated learning pipeline across Singapore's major hospital clusters that trains a shared model without centralising patient data. Each cluster trains locally; only model weights are shared.
- Pathology-specific augmentation — For oral health (SMILE AI), we created augmentation pipelines that simulate common Southeast Asian oral pathologies not well-represented in global datasets.
- Radiologist-in-the-loop annotation — Budget for at least 2 radiologist reviews per case for training data. Disagreements between annotators are often the most informative training signal.
Key takeaways
- Models trained on Western cohorts often underperform on Singapore and regional patient populations.
- A modest local dataset frequently beats massive transfer learning from foreign benchmarks.
- Federated learning enables multi-cluster training without centralising patient data.
- Pathology-specific augmentation matters for conditions under-represented globally.
- Budget for multi-radiologist annotation — disagreement is valuable training signal.
FAQ
Why doesn't transfer learning from global imaging datasets suffice in Singapore?
Anatomy, disease prevalence, and imaging protocols differ; local fine-tuning on even hundreds of cases can close double-digit performance gaps.
How can hospitals collaborate without sharing raw patient images?
Federated pipelines train locally at each cluster and exchange model weights only, preserving PDPA-aligned data boundaries.
What annotation standard should imaging AI projects plan for?
At least two radiologist reviews per training case; resolve or preserve disagreement explicitly rather than forcing false consensus.