Construction and Validation of a General Medical Image Dataset for Pretraining

Area of Science:

• Deep learning
• Medical image analysis
• Computer vision

Background:

• Current deep learning models for medical imaging often train from scratch or use ImageNet pretraining.
• A universally accepted medical image dataset for pretraining is lacking.
• Transfer learning enhances model performance but requires suitable pretraining datasets.

Purpose of the Study:

• To construct a general-purpose medical image dataset for model pretraining.
• To validate the effectiveness of models pretrained on this dataset for downstream tasks.
• To establish a new standard for pretraining in medical image analysis.

Main Methods:

• Collected and curated several public medical image datasets to form the CPMID dataset.
• Pretrained Resnet and Vision Transformer models using the CPMID dataset.
• Evaluated pretrained models on medical image classification and segmentation tasks, comparing against scratch and ImageNet pretraining.

Main Results:

• Models pretrained on CPMID achieved superior classification accuracy and ROC-AUC values compared to ImageNet-pretrained models.
• Average classification accuracy improvements of 4.30%, 8.86%, and 3.85% were observed.
• Optimal balance of performance and efficiency was demonstrated in both classification and segmentation tasks.

Conclusions:

• The constructed CPMID dataset and its associated pretrained models are highly effective for medical image analysis.
• CPMID-based pretraining offers significant advantages over ImageNet pretraining for medical imaging tasks.
• This work provides a valuable resource for advancing deep learning in medical image analysis.