Deep Learning for Fast and Spatially-Constrained Tissue Quantification from Highly-Undersampled Data in Magnetic...

Area of Science:

• Medical Imaging
• Quantitative MRI
• Biophysics

Background:

• Magnetic Resonance Fingerprinting (MRF) enables simultaneous measurement of tissue properties like T1 and T2 relaxation times.
• Current MRF methods, while efficient, require further acceleration, particularly for pediatric imaging.
• Conventional MRF lacks spatial context in its template matching algorithm, limiting quantification accuracy.

Purpose of the Study:

• To develop an accelerated MRF post-processing technique using fewer sampling data.
• To enhance quantification accuracy by incorporating spatial information from surrounding pixels.
• To accelerate MRF acquisition for improved clinical applicability, especially in vulnerable populations.

Main Methods:

• A novel post-processing method was developed to accelerate MRF acquisition.
• A deep learning model, U-Net, was employed to map MRF signal evolutions to tissue property maps.
• Principal Component Analysis (PCA) was used to reduce the dimensionality of input MRF signals for U-Net.

Main Results:

• The proposed method accurately quantifies T1 and T2 relaxation times using only 25% of the original time points.
• This represents a four-fold acceleration in MRF data acquisition compared to conventional template matching.
• Validation was performed using in vivo human brain data.

Conclusions:

• The developed deep learning approach significantly accelerates MRF acquisition while preserving quantification accuracy.
• Incorporating spatial information and deep learning improves upon traditional MRF methods.
• This advancement holds promise for faster and more efficient quantitative MRI, especially in pediatric applications.