Targeted transfer learning to improve performance in small medical physics datasets.

Area of Science:

• Medical imaging
• Machine learning
• Radiology

Background:

• Deep learning (DL) models show promise in medical image analysis.
• Training DL models on limited medical datasets presents significant challenges.
• Optimizing DL model performance requires careful consideration of training strategies.

Purpose of the Study:

• To evaluate state-of-the-art techniques for training neural networks on small medical imaging datasets.
• To identify optimal approaches for deep learning model development in resource-limited scenarios.
• To assess the impact of transfer learning strategies on model performance with varying dataset sizes.

Main Methods:

• Utilized 112,120 frontal-view X-ray images from the NIH ChestXray14 dataset.
• Trained DensNet121 and ResNet50 convolutional neural networks (CNNs).
• Investigated transfer learning from ImageNet and X-ray datasets, layer freezing strategies, learning rate policies, and data quantity effects (N=50 to 77,880).

Main Results:

• For small datasets (<2000 samples), performance significantly improved (>15% AUC) with optimized transfer learning, dropout, cyclic learning rates, and dynamic layer freezing.
• Sufficient data (>35,000 samples) minimized the need for extensive parameter tuning.
• Transfer learning using same-anatomical-site X-ray images yielded the best results, outperforming ImageNet pretraining by up to 15%, enabling training with as few as 50 samples.

Conclusions:

• Training deep learning models on small datasets (<2000) is feasible with optimized transfer learning.
• Large datasets (>35,000) require minimal tuning for high performance.
• Transfer learning with same-site X-ray images is highly effective, even with minimal data (N=50), suggesting learned features may be less general than previously assumed.