LAUSR

Editorial for “Image Quality Assessment of Fetal Brain MRI Using Multi-Instance Deep Learning Methods” Ultrasound (US) is currently the predominant modality for primary assessment of fetal anatomy and maternal conditions during pregnancy; however, fetal magnetic resonance imaging (MRI) has been increasingly used in clinical practice and is considered as a complementary technology, which is not affected by maternal obesity, oligohydramnios, multiple fetuses, fetal position, and bones, and can offer additional anatomical and pathological information that cannot be accurately provided by US. With its larger field of view, higher spatial resolution, better image contrast, as well as nonionizing radiation, fetal MRI has the unique ability to detect subtle abnormalities in the brain, facilitates the evaluation of fetuses with large or complex anomalies, and visualization of lesions within the whole-body of the fetus. Qualified image quality is the priority of accurate qualitative diagnosis and exact quantitative measurement in both clinical and research practice. Owing to random motion of the fetuses and maternal respiration, fetal brain MRI is confronted with great challenges to achieve good image quality that satisfies diagnosis. The implications of mis-, over-, and underdiagnoses can be significant; all images of the whole imaging protocol should be assessed before the final report because the findings can be hiding in motion degraded pulse sequences. Recently, rapid MRI sequences and new threedimensional reconstruction algorithms have been developed and used to overcome these issues. Ultrafast pulse sequences, such as single-shot fast spin echo (SSFSE), half-Fourier acquisition single-shot turbo spin echo (HASTE), echo-planar imaging (EPI), and innervolume echo volumar imaging (IVEVI), have been used for fetal brain MRI. In addition, novel image reconstruction methods, including sparse volume reconstruction based on adaptive kernel regression as well as reconstruction with fully automatic framework, are proposed to improve algorithm efficiency and image quality. Although those sequences and reconstruction algorithms have improved the overall image quality, it still needs to visually assess the images, which is carried out within the imaging phase and after data reconstruction. In recent years, automatic methods based on deep learning methods (DLMs) have been used to rapidly and objectively evaluate the image quality compared with conventional visual inspection. In this issue of the Journal of Magnetic Resonance Imaging, Largent et al proposed, assessed and compared three multi-instance deep learning methods (MI DLMs), MI count-based DLM (MI-CB-DLM), MI vote-based DLM (MI-VB-DLM), and MI feature embedding DLM (MI-FEDLM), for automatic assessment of fetal brain MR image quality after three-dimensional reconstruction as well as to explore the influence of fetal gestational age (GA) on the performance of the MI DLMs. Two hundred and seventy-one MR scans were performed for 211 fetuses and acquired in orthorhombic (axial, sagittal, and coronal) planes using a T2-weighted two-dimensional single-shot fast spin-echo sequence. It was found that MI-CB-DLM outperformed the other two MI DLMs in assessing the image quality of T2-weighted fetal brain images after three-dimensional reconstruction, which further improved its performance when including GA as an input variable of this MI DLM. The current finding shows that MI-CB-DLM may potentially serve as a good technique to objectively and rapidly evaluate fetal MR image quality, which also further strength automatic image quality assessment using DLM. However, it is a retrospective study in one study center and single MRI system, lacking external validation. Moreover, the acquisition time (3 minutes) per SSFSE scan was a bit long in Largent’s study, which also induced motion artifacts due to the longish scan time. Accelerated imaging techniques, simultaneous multislice, compressed-sensing, deep learning-based acceleration, can be developed and integrated with the existing imaging protocol to decrease the scan time. In short, the authors’ findings improve automatic assessment of image quality as well as advance the field of fetal brain MRI. Despite the currently reported DLMs have shown great potential and advantages in assessing image quality of fetal brain MRI, which still needs to enlarge sample size, enrolls more imaging centers with different MR vendors, and refines DLMs in further studies.