Overcoming data scarcity in biomedical imaging with a foundational multi-task model.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Biomedical Informatics

Background:

• Foundational models show promise but require large datasets, which are scarce in biomedical imaging.
• Existing pretraining methods struggle with the specialized nature of medical data.

Purpose of the Study:

• To develop a universal biomedical pretrained model (UMedPT) using a multi-task learning strategy.
• To address memory constraints and data limitations in medical imaging AI.

Main Methods:

• Trained UMedPT on a diverse multi-task database including tomographic, microscopic, and X-ray images.
• Employed various labeling strategies: classification, segmentation, and object detection.
• Utilized a multi-task learning approach to decouple training tasks from memory needs.

Main Results:

• UMedPT outperformed ImageNet pretraining and state-of-the-art models on medical imaging tasks.
• Achieved high performance with only 1% of training data for in-domain tasks and 50% for out-of-domain tasks, without fine-tuning.
• Demonstrated superior cross-center transferability in external validation.

Conclusions:

• UMedPT offers a highly efficient and effective foundational model for biomedical imaging.
• The multi-task learning strategy overcomes data limitations and improves model generalizability.
• UMedPT sets a new benchmark for feature extraction and transferability in medical AI.