Self-supervised and transfer learning make deep learning denoising practical for accelerated brain MRI

Deep learning denoising of accelerated 3D T1-weighted brain MRI with transfer-learning fine-tuning or self-supervised Self2Self-AM achieved structural similarity index 0.943, cortical thickness discrepancy 0.16 mm, Dice 0.986, beating conventional denoising.

• Self2Self-AM trains the convolutional neural network on the noisy image volume itself, removing the need for additional high-signal-to-noise reference data.
• Transfer learning enabled supervised denoising using high-signal-to-noise data from a single subject and substantially reduced Self2Self-AM training time.
• Both approaches significantly improved brain morphometric quantification over raw images and conventional denoising benchmarks.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.