Deep learning CAIPIRINHA-accelerated 3D MRI of the knee cartilage at 7 Tesla

Abstract

Purpose

To investigate the feasibility and degree of scan-time reduction that can be achieved by deep learning (DL)-reconstruction in 3D dual-echo steady-state (DESS) knee-cartilage imaging accelerated with CAIPIRINHA (controlled aliasing in parallel imaging results in higher acceleration) at 7 Tesla (T).

Methods

35 knee joints of 18 volunteers were imaged using an isotropic fat-suppressed DL-reconstructed CAIPIRINHA-accelerated DESS sequence (CAIPIRINHA_DL) with an isotropic resolution of 0.25 mm. CAIPIRINHA_DL acquisitions were accelerated four-, six-, eight-, and twelve-fold (CAIPIRINHA_DL4-CAIPIRINHA_DL12), and compared to a state-of-the-art, fourfold accelerated GRAPPA (generalized autocalibrating partially parallel acquisitions) DESS acquisition (GRAPPA₄) with matched spatial resolution. Three readers rated image-quality criteria. Rating differences were characterized with Friedman’s analysis of variance. Inter-reader agreement was assessed for grading of incidental cartilage lesions.

Results

CAIPIRINHA_DL6 had the shortest acquisition time (4:04 min) among sequences whose images were rated in the best category, with no statistically significant difference (p > 0.999) to GRAPPA₄ (5:53 min). The overall quality of CAIPIRINHA_DL8 and CAIPIRINHA_DL12 images was rated lower than GRAPPA₄ images, while artifacts were rated significantly more severe (p < 0.001, each). Inter-reader agreement for cartilage lesion grading ranged from almost perfect for CAIPIRINHA_DL4 and GRAPPA₄ (κ = 0.90, each) to moderate for CAIPIRINHA_DL12 (κ = 0.52).

Conclusions

The quality of DL-reconstructed 3D DESS images acquired at 7 T with four- and six-fold CAIPIRINHA-based acceleration was equivalent to that of 3D-DESS images acquired with four-fold GRAPPA-based acceleration. These findings suggest the feasibility of DL-reconstructions for imaging of the knee cartilage at 7 T field strength.