Clean Self-Supervised MRI Reconstruction from Noisy, Sub-Sampled Training Data with Robust SSDU.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Fully supervised deep learning for MRI reconstruction requires impractical fully sampled, high SNR datasets.
• Existing self-supervised methods struggle with noise in sub-sampled training data.

Purpose of the Study:

• To develop robust self-supervised deep learning methods for MRI reconstruction using noisy, sub-sampled data.
• To improve image quality and reduce reconstruction errors in MRI.

Main Methods:

• Proposed Robust SSDU, which simultaneously estimates missing k-space samples and denoises data.
• Robust SSDU trains networks using a noisy-to-less-noisy mapping with a Noisier2Noise correction.
• Introduced Noiser2Full for reconstruction from noisy, fully sampled data.

Main Results:

• Robust SSDU provably recovers clean images from noisy, sub-sampled training data.
• Methods are architecture-agnostic, easy to implement, and computationally similar to standard training.
• Evaluated on fastMRI brain dataset, achieving competitive performance against a clean data benchmark.

Conclusions:

• Robust SSDU and Noiser2Full offer effective solutions for MRI reconstruction with limited or noisy data.
• These methods advance self-supervised learning in medical imaging, reducing reliance on ideal training datasets.