ENSURE: ENSEMBLE STEIN'S UNBIASED RISK ESTIMATOR FOR UNSUPERVISED LEARNING

Area of Science:

• Medical imaging
• Artificial intelligence
• Signal processing

Background:

• Deep learning offers advantages over compressed sensing for image reconstruction.
• Acquiring fully sampled training images is often infeasible in medical imaging applications.
• Existing methods using Stein's Unbiased Risk Estimator (SURE) for training deep networks face challenges with mean square error approximation in undersampled settings.

Purpose of the Study:

• To develop a novel deep learning training approach that circumvents the need for fully sampled images.
• To enable effective end-to-end training of deep networks using only undersampled measurements.
• To improve image reconstruction quality in scenarios where complete data is unavailable.

Main Methods:

• Proposed an ENsemble SURE (ENSURE) method for training deep networks.
• Utilized an ensemble of images, each with a different undersampled acquisition pattern.
• Demonstrated that this ensemble approach approximates the mean square error (MSE) effectively.

Main Results:

• The ENSURE approach enables training deep networks solely from undersampled data.
• Preliminary results show comparable reconstruction quality to supervised learning methods.
• The method achieves similar performance to recent unsupervised learning techniques.

Conclusions:

• The ENSURE method provides a viable solution for training deep learning models in undersampled medical imaging.
• This approach overcomes limitations of previous SURE-based training strategies.
• ENSURE facilitates high-quality image reconstruction without requiring fully sampled training data.