Technical note: Progressive deep learning: An accelerated training strategy for medical image segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep Learning (DL) excels in medical image analysis but demands extensive computation and time.
• Training DL models involves complex networks and large datasets, leading to lengthy processes.
• Manual hyperparameter optimization for DL networks is intensive and repetitive.

Purpose of the Study:

• To present Progressive Deep Learning (PDL), a new approach for accelerating DL model training.
• To reduce training time for medical image segmentation using a progressive dataset feeding strategy.
• To maintain segmentation performance while enhancing training efficiency.

Main Methods:

• A two-stage PDL approach was applied to CT and MRI auto-segmentation tasks.
• Training datasets were ranked using similarity metrics (MSE, PSNR, SSIM, UQI).
• Coarse sampling of high-ranked datasets optimized hyperparameters, followed by full dataset integration for faster convergence.

Main Results:

• PDL reduced training time by approximately 49% across CT and MRI segmentation.
• Segmentation performance (Dice coefficient) remained statistically similar to conventional DL methods.
• Significant reductions in training duration were observed for U-Net and DenseNet architectures on both modalities.

Conclusions:

• The PDL approach substantially decreases training time for medical image segmentation.
• PDL maintains high segmentation performance comparable to traditional DL training methods.
• This method offers an efficient solution for training deep learning models in medical imaging.