Related Experiment Video
Updated: May 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
Large-scale benchmarking and boosting transfer learning for medical image analysis
Mohammad Reza Hosseinzadeh Taher1, Fatemeh Haghighi1, Michael B Gotway2
1School of Computing and Augmented Intelligence, Arizona State University, Tempe, AZ 85281, USA.
Medical Image Analysis
|March 21, 2025
Summary
Convolutional Neural Networks (ConvNets) outperform vision transformers for medical image analysis transfer learning. ConvNets are more annotation-efficient, and fine-grained features are crucial for specific medical tasks.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computer Vision
Background:
- Transfer learning is vital for medical image analysis, but selecting optimal pretrained models is challenging due to numerous architectures and pretraining strategies.
- A lack of large-scale, up-to-date evaluations hinders practitioners in choosing the best models for their specific medical imaging tasks.
Purpose of the Study:
- To systematically evaluate the transferability of various deep learning models, including ConvNets and vision transformers, to diverse medical imaging tasks.
- To investigate the impact of fine-tuning data size, pretraining data granularity, and self-supervised learning methods on transfer learning performance in medicine.
- To assess the effectiveness of domain-adaptive pretraining for developing high-performance medical imaging models.
Main Methods:
- Benchmarking numerous conventional and modern ConvNet and vision transformer architectures across various medical tasks.
- Investigating the influence of fine-tuning dataset size and pretraining data granularity on model performance.
- Evaluating a wide range of recent self-supervised learning methods and domain-adaptive pretraining strategies.
- Conducting approximately 5,000 experiments to ensure comprehensive analysis.
Main Results:
- ConvNets exhibit superior transferability and annotation efficiency compared to vision transformers in medical imaging tasks.
- Fine-grained representations are more critical than high-level semantic features for fine-grained medical image analysis.
- Self-supervised models demonstrate advantages in learning holistic features over supervised models.
- Domain-adaptive pretraining effectively enhances model performance by leveraging both general (ImageNet) and medical-specific datasets.
Conclusions:
- ConvNets remain a robust choice for transfer learning in medical imaging, particularly when annotation resources are limited.
- The choice of pretraining strategy, data granularity, and model architecture significantly impacts performance on specific medical tasks.
- Domain-adaptive pretraining offers a promising avenue for developing highly effective medical image analysis tools.
