RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning.

Area of Science:

• Medical Imaging AI
• Transfer Learning in Medicine
• Radiologic Image Analysis

Background:

• Deep learning models for medical imaging often struggle with limited annotated datasets.
• Pretraining on large, diverse datasets is crucial for improving model generalization.
• ImageNet, a dataset of photographic images, is commonly used for pretraining, but may not be optimal for medical tasks.

Purpose of the Study:

• To evaluate the efficacy of pretraining models on a large dataset of radiologic images (RadImageNet) versus photographic images (ImageNet) for downstream medical applications.
• To compare the performance of RadImageNet and ImageNet pretrained models in classification and segmentation tasks using transfer learning.

Main Methods:

• A retrospective study extracted 1.35 million annotated medical images from CT, MRI, and US studies.
• Models were pretrained on the RadImageNet dataset and compared against models pretrained on ImageNet.
• Performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUC) for classification and Dice scores for segmentation tasks.

Main Results:

• RadImageNet models significantly outperformed ImageNet models across eight classification tasks, showing AUC improvements ranging from 0.9% to 19.4% (P < .001).
• Notable improvements were observed in smaller datasets, such as thyroid nodules (9.4% AUC) and breast masses (4.0% AUC).
• Lesion localization accuracy also improved significantly on ultrasound datasets (64.6% for thyroid, 16.4% for breast).

Conclusions:

• Pretraining with RadImageNet offers a substantial advantage over ImageNet for medical transfer learning tasks, especially with limited data.
• RadImageNet models demonstrate enhanced interpretability compared to ImageNet models in radiologic applications.
• The findings underscore the value of using domain-specific datasets for pretraining medical AI models.