Deep learning-based methods may minimize GBCA dosage in brain MRI

Abstract

Objectives: To evaluate the clinical performance of a deep learning (DL)-based method for brain MRI exams with reduced gadolinium-based contrast agent (GBCA) dose to provide better understanding of the readiness and limitations of this method.

Methods: Eighty-three consecutive patients (from March 2019 to August 2019) who underwent brain contrast-enhanced (CE) MRI were included. Three 3D T1-weighted images with zero-dose, low-dose (10%), and full-dose (100%) GBCA were collected. The first 30 cases were used to train a DL model to synthesize the full-dose GBCA images from the zero-dose and low-dose image pairs. The remaining 53 cases were used for testing. The enhancement pattern, number, and location of enhancing lesions were recorded. Overall image quality, image signal noise ratio (SNR), lesion conspicuity, and lesion enhancement were assessed.

Results: Lesion detection from the DL-synthesized CE-MRI image accurately matched those from the true full-dose CE-MRI images in 48 of 53 cases (90.6%). The DL method identified the lesions in 34 of 36 cases (94.4%) with a single enhanced lesion and all lesions in 3 of 6 cases (50.0%) in cases with multiple enhancing lesions. The agreement between synthesized and true full-dose CE-MRI images were 0.73, 0.63, 0.89, and 0.87 for image quality, image SNR, lesion conspicuity, and lesion enhancement, respectively.

Conclusions: The proposed DL method is a feasible way to minimize the dosage of GBCAs in brain MRI without sacrificing the diagnostic information. Missing enhancement of small lesions in patients with multiple lesions was observed, requiring improvements in algorithms or dosage design.

Key points: • This study evaluated the clinical performance of a DL-based reconstruction method for significant dose reduction in GBCA contrast-enhanced MRI exams. • The proposed DL method has the potential to satisfy the routine radiological diagnosis needs in certain clinical applications.

Keywords: Contrast media; Deep learning; Gadolinium; Magnetic resonance imaging.