A Systematic Benchmarking Analysis of Transfer Learning for Medical Image Analysis

Area of Science:

• Deep learning in medical imaging
• Computer vision applications in healthcare
• Transfer learning methodologies

Background:

• Supervised transfer learning from ImageNet is common in medical image analysis.
• A large-scale benchmark for new pre-training techniques in medical imaging is lacking.
• Existing methods do not fully address the domain gap between natural and medical images.

Purpose of the Study:

• To systematically evaluate transfer learning from diverse pre-trained models (iNat2021, self-supervised ImageNet) on medical tasks.
• To compare their efficacy against supervised ImageNet models.
• To introduce and assess a continual pre-training approach to bridge the domain gap.

Main Methods:

• Systematic evaluation of 14 self-supervised ImageNet models and iNat2021 models.
• Benchmarking on 7 diverse medical imaging tasks.
• Implementation of continual pre-training on medical images.

Main Results:

• Models pre-trained on fine-grained data (iNat2021) provide superior local representations for medical segmentation.
• Self-supervised ImageNet models demonstrate more effective learning of holistic features compared to supervised models.
• Continual pre-training successfully mitigates the domain gap between natural and medical image datasets.

Conclusions:

• Pre-training strategies significantly impact performance on specific medical imaging tasks.
• Self-supervised and fine-grained pre-training offer distinct advantages over traditional supervised ImageNet transfer learning.
• Continual pre-training is a viable strategy for enhancing transfer learning in medical AI.