We present a critical assessment of the role of transfer learning in training fully convolutional networks (FCNs) for medical image segmentation. We first show that although transfer learning reduces the training time on the target task, improvements in segmentation accuracy are highly task/data-dependent. Large improvements are observed only when the segmentation task is more challenging and the target training data is smaller. We shed light on these observations by investigating the impact of transfer learning on the evolution of model parameters and learned representations. We observe that convolutional filters change little during training and still look random at convergence. We further show that quite accurate FCNs can be built by freezing the encoder section of the network at random values and only training the decoder section. At least for medical image segmentation, this finding challenges the common belief that the encoder section needs to learn data/task-specific representations. We examine the evolution of FCN representations to gain a deeper insight into the effects of transfer learning on the training dynamics. Our analysis shows that although FCNs trained via transfer learning learn different representations than FCNs trained with random initialization, the variability among FCNs trained via transfer learning can be as high as that among FCNs trained with random initialization. Moreover, feature reuse is not restricted to the early encoder layers; rather, it can be more significant in deeper layers. These findings offer new insights and suggest alternative ways of training FCNs for medical image segmentation.

The relationship between structural changes of the cerebral cortex revealed by Magnetic Resonance Imaging (MRI) and gene expression in the human fetal brain has not been explored. In this study, we aimed to test the hypothesis that relative regional thickness (a measure of cortical evolving organization) of fetal cortical compartments (cortical plate [CP] and subplate [SP]) is associated with expression levels of genes with known cortical phenotype. Mean regional SP/CP thickness ratios across age measured on in utero MRI of 25 healthy fetuses (20-33 gestational weeks [GWs]) were correlated with publicly available regional gene expression levels (23-24 GW fetuses). Larger SP/CP thickness ratios (more pronounced cortical evolving organization) was found in perisylvian regions. Furthermore, we found a significant association between SP/CP thickness ratio and expression levels of the FLNA gene (mutated in periventricular heterotopia, congenital heart disease, and vascular malformations). Further work is needed to identify early MRI biomarkers of gene expression that lead to abnormal cortical development.

OBJECTIVE: Approximately 50% of patients with tuberous sclerosis complex develop infantile spasms, a sudden onset epilepsy syndrome associated with poor neurological outcomes. An increased burden of tubers confers an elevated risk of infantile spasms, but it remains unknown whether some tuber locations confer higher risk than others. Here, we test whether tuber location and connectivity are associated with infantile spasms. METHODS: We segmented tubers from 123 children with (n = 74) and without (n = 49) infantile spasms from a prospective observational cohort. We used voxelwise lesion symptom mapping to test for an association between spasms and tuber location. We then used lesion network mapping to test for an association between spasms and connectivity with tuber locations. Finally, we tested the discriminability of identified associations with logistic regression and cross-validation as well as statistical mediation. RESULTS: Tuber locations associated with infantile spasms were heterogenous, and no single location was significantly associated with spasms. However, >95% of tuber locations associated with spasms were functionally connected to the globi pallidi and cerebellar vermis. These connections were specific compared to tubers in patients without spasms. Logistic regression found that globus pallidus connectivity was a stronger predictor of spasms (odds ratio [OR] = 1.96, 95% confidence interval [CI] = 1.10-3.50, p = 0.02) than tuber burden (OR = 1.65, 95% CI = 0.90-3.04, p = 0.11), with a mean receiver operating characteristic area under the curve of 0.73 (±0.1) during repeated cross-validation. INTERPRETATION: Connectivity between tuber locations and the bilateral globi pallidi is associated with infantile spasms. Our findings lend insight into spasm pathophysiology and may identify patients at risk. ANN NEUROL 2021;89:726-739.

BACKGROUND AND PURPOSE: Little is known about microstructural development of cerebellar white matter in vivo. This study aimed to investigate developmental changes of the cerebellar peduncles in second- and third-trimester healthy fetuses using motion-corrected DTI and tractography. MATERIALS AND METHODS: 3T data of 81 healthy fetuses were reviewed. Structural imaging consisted of multiplanar T2-single-shot sequences; DTI consisted of a series of 12-direction diffusion. A robust motion-tracked section-to-volume registration algorithm reconstructed images. ROI-based deterministic tractography was performed using anatomic landmarks described in postnatal tractography. Asymmetry was evaluated qualitatively with a perceived difference of >25% between sides. Linear regression evaluated gestational age as a predictor of tract volume, ADC, and fractional anisotropy. RESULTS: Twenty-four cases were excluded due to low-quality reconstructions. Fifty-eight fetuses with a median gestational age of 30.6 weeks (interquartile range, 7 weeks) were analyzed. The superior cerebellar peduncle was identified in 39 subjects (69%), and it was symmetric in 15 (38%). The middle cerebellar peduncle was identified in all subjects and appeared symmetric; in 13 subjects (22%), two distinct subcomponents were identified. The inferior cerebellar peduncle was not found in any subject. There was a significant increase in volume for the superior cerebellar peduncle and middle cerebellar peduncle (both,

BACKGROUND: Diffusion-weighted MRI (DW-MRI) of the kidneys is a technique that provides information about the microstructure of renal tissue without requiring exogenous contrasts such as gadolinium, and it can be used for diagnosis in cases of renal disease and assessing response-to-therapy. However, physiological motion and large geometric distortions due to main B0 field inhomogeneities degrade the image quality, reduce the accuracy of quantitative imaging markers, and impede their subsequent clinical applicability. PURPOSE: To retrospectively correct for geometric distortion for free-breathing DW-MRI of the kidneys at 3T, in the presence of a nonstatic distortion field due to breathing and bulk motion. STUDY TYPE: Prospective. SUBJECTS: Ten healthy volunteers (ages 29-38, four females). FIELD STRENGTH/SEQUENCE: 3T; DW-MR dual-echo echo-planar imaging (EPI) sequence (10 b-values and 17 directions) and a T2 volume. ASSESSMENT: The distortion correction was evaluated subjectively (Likert scale 0-5) and numerically with cross-correlation between the DW images at b = 0 s/mm2 and a T2 volume. The intravoxel incoherent motion (IVIM) and diffusion tensor (DTI) model-fitting performance was evaluated using the root-mean-squared error (nRMSE) and the coefficient of variation (CV%) of their parameters. STATISTICAL TESTS: Statistical comparisons were done using Wilcoxon tests. RESULTS: The proposed method improved the Likert scores by 1.1 ± 0.8 (P < 0.05), the cross-correlation with the T2 reference image by 0.13 ± 0.05 (P < 0.05), and reduced the nRMSE by 0.13 ± 0.03 (P < 0.05) and 0.23 ± 0.06 (P < 0.05) for IVIM and DTI, respectively. The CV% of the IVIM parameters (slow and fast diffusion, and diffusion fraction for IVIM and mean diffusivity, and fractional anisotropy for DTI) was reduced by 2.26 ± 3.98% (P = 6.971 × 10-2 ), 11.24 ± 26.26% (P = 6.971 × 10-2 ), 4.12 ± 12.91% (P = 0.101), 3.22 ± 0.55% (P < 0.05), and 2.42 ± 1.15% (P < 0.05). DATA CONCLUSION: The results indicate that the proposed Di + MoCo method can effectively correct for time-varying geometric distortions and for misalignments due to breathing motion. Consequently, the image quality and precision of the DW-MRI model parameters improved. LEVEL OF EVIDENCE: 2 TECHNICAL EFFICACY STAGE: 1.

Fetal cortical plate segmentation is essential in quantitative analysis of fetal brain maturation and cortical folding. Manual segmentation of the cortical plate, or manual refinement of automatic segmentations is tedious and time-consuming. Automatic segmentation of the cortical plate, on the other hand, is challenged by the relatively low resolution of the reconstructed fetal brain MRI scans compared to the thin structure of the cortical plate, partial voluming, and the wide range of variations in the morphology of the cortical plate as the brain matures during gestation. To reduce the burden of manual refinement of segmentations, we have developed a new and powerful deep learning segmentation method. Our method exploits new deep attentive modules with mixed kernel convolutions within a fully convolutional neural network architecture that utilizes deep supervision and residual connections. We evaluated our method quantitatively based on several performance measures and expert evaluations. Results show that our method outperforms several state-of-the-art deep models for segmentation, as well as a state-of-the-art multi-atlas segmentation technique. We achieved average Dice similarity coefficient of 0.87, average Hausdorff distance of 0.96 mm, and average symmetric surface difference of 0.28 mm on reconstructed fetal brain MRI scans of fetuses scanned in the gestational age range of 16 to 39 weeks (28.6± 5.3). With a computation time of less than 1 minute per fetal brain, our method can facilitate and accelerate large-scale studies on normal and altered fetal brain cortical maturation and folding.

OBJECTIVE: Congenital heart disease (CHD) is associated with abnormal brain development in utero. We applied innovative fetal magnetic resonance imaging (MRI) techniques to determine whether reduced fetal cerebral substrate delivery impacts the brain globally, or in a region-specific pattern. Our novel design included two control groups, one with and the other without a family history of CHD, to explore the contribution of shared genes and/or fetal environment to brain development. METHODS: From 2014 to 2018, we enrolled 179 pregnant women into 4 groups: "HLHS/TGA" fetuses with hypoplastic left heart syndrome (HLHS) or transposition of the great arteries (TGA), diagnoses with lowest fetal cerebral substrate delivery; "CHD-other," with other CHD diagnoses; "CHD-related," healthy with a CHD family history; and "optimal control," healthy without a family history. Two MRIs were obtained between 18 and 40 weeks gestation. Random effect regression models assessed group differences in brain volumes and relationships to hemodynamic variables. RESULTS: HLHS/TGA (n = 24), CHD-other (50), and CHD-related (34) groups each had generally smaller brain volumes than the optimal controls (71). Compared with CHD-related, the HLHS/TGA group had smaller subplate (-13.3% [standard error = 4.3%], p

In MRI practice, it is inevitable to appropriately balance between image resolution, signal-to-noise ratio (SNR), and scan time. It has been shown that super-resolution reconstruction (SRR) is effective to achieve such a balance, and has obtained better results than direct high-resolution (HR) acquisition, for certain contrasts and sequences. The focus of this work was on constructing images with spatial resolution higher than can be practically obtained by direct Fourier encoding. A novel learning approach was developed, which was able to provide an estimate of the spatial gradient prior from the low-resolution (LR) inputs for the HR reconstruction. By incorporating the anisotropic acquisition schemes, the learning model was trained over the LR images themselves only. The learned gradients were integrated as prior knowledge into a gradient-guided SRR model. A closed-form solution to the SRR model was developed to obtain the HR reconstruction. Our approach was assessed on the simulated data as well as the data acquired on a Siemens 3T MRI scanner containing 45 MRI scans from 15 subjects. The experimental results demonstrated that our approach led to superior SRR over state-of-the-art methods, and obtained better images at lower or the same cost in scan time than direct HR acquisition.

Supervised training of deep neural networks in medical imaging applications relies heavily on expert-provided annotations. These annotations, however, are often imperfect, as voxel-by-voxel labeling of structures on 3D images is difficult and laborious. In this paper, we focus on one common type of label imperfection, namely, false negatives. Focusing on brain lesion detection, we propose a method to train a convolutional neural network (CNN) to segment lesions while simultaneously improving the quality of the training labels by identifying false negatives and adding them to the training labels. To identify lesions missed by annotators in the training data, our method makes use of the 1) CNN predictions, 2) prediction uncertainty estimated during training, and 3) prior knowledge about lesion size and features. On a dataset of 165 scans of children with tuberous sclerosis complex from five centers, our method achieved better lesion detection and segmentation accuracy than the baseline CNN trained on the noisy labels, and than several alternative techniques.

Functional MRI (fMRI) is extremely challenging to perform in subjects who move because subject motion disrupts blood oxygenation level dependent (BOLD) signal measurement. It has become common to use retrospective framewise motion detection and censoring in fMRI studies to eliminate artifacts arising from motion. Data censoring results in significant loss of data and statistical power unless the data acquisition is extended to acquire more data not corrupted by motion. Acquiring more data than is necessary leads to longer than necessary scan duration, which is more expensive and may lead to additional subject non-compliance. Therefore, it is well established that real-time prospective motion monitoring is crucial to ensure data quality and reduce imaging costs. In addition, real-time monitoring of motion allows for feedback to the operator and the subject during the acquisition, to enable intervention to reduce the subject motion. The most widely used form of motion monitoring for fMRI is based on volume-to-volume registration (VVR), which quantifies motion as the misalignment between subsequent volumes. However, motion is not constrained to occur only at the boundaries of volume acquisition, but instead may occur at any time. Consequently, each slice of an fMRI acquisition may be displaced by motion, and assessment of whole volume to volume motion may be insensitive to both intra-volume and inter-volume motion that is revealed by displacement of the slices. We developed the first slice-by-slice self-navigated motion monitoring system for fMRI by developing a real-time slice-to-volume registration (SVR) algorithm. Our real-time SVR algorithm, which is the core of the system, uses a local image patch-based matching criterion along with a Levenberg-Marquardt optimizer, all accelerated via symmetric multi-processing, with interleaved and simultaneous multi-slice acquisition schemes. Extensive experimental results on real motion data demonstrated that our fast motion monitoring system, named Slice Localization Integrated MRI Monitoring (SLIMM), provides more accurate motion measurements than a VVR based approach. Therefore, SLIMM offers improved online motion monitoring which is particularly important in fMRI for challenging patient populations. Real-time motion monitoring is crucial for online data quality control and assurance, for enabling feedback to the subject and the operator to act to mitigate motion, and in adaptive acquisition strategies that aim to ensure enough data of sufficient quality is acquired without acquiring excess data.