RPS 1013 - Advances in MRI

Purpose:

Describe how slice-to-slice displacement occurring in 2D multi-slice acquisition of overlapped slices can be corrected, allowing high quality reformats in other orientations.

Super-resolution is a method which provides improved through-plane resolution in a 2D acquisition by using input data comprised of overlapped slices. T2-weighted spin-echo (T2SE) imaging is typically acquired in multiple passes to avoid slice-to-slice interference. The resultant e.g. axial images can be reformatted into other orientations, e.g. sagittal or coronal. For static objects, this can be effective. However, for subtle (≈1 mm) motion as in T2SE of the prostate, the reformats can have objectionable “scalloping” artefact. Here we describe a correction algorithm. Because consecutive axial slices are overlapped, they are highly correlated. Using cross-correlation we determine and correct for slice-to-slice displacement between adjacent slices. The knowledge that subsets of slices are acquired within the same pass allows additional constraints on the correction. The method was assessed in 16 patients undergoing prostate MRI. Each was imaged using 78 3-mm thick slices with 2 mm slice-to-slice overlap. Slice sets with and without motion correction were subjected to super-resolution processing, and sagittal reformats were evaluated for artefact inhibiting visualisation of the margin between the prostate and anterior rectal wall.

In 13 of the 16 studies, the level of the artifact was reduced modestly or significantly. Plots displacements, used for correction, from a fiducial (X, Y) mean value in the transverse direction of all 78 slices acquired. Figure 2 compares sagittal reformats using uncorrected (left) and corrected (right) axial image sets.

Retrospective motion correction of T2SE axial images of the prostate allows high-resolution artefact-suppressed reformatting in other orientations.

Limited number of studies.

This study was approved by the IRB.

This work was funded (NIH RR018898).

Purpose:

The aim of the study was to compensate for MR field non-uniformities of breast MRI on retrospective data sets, a prerequisite step to compute robust radiomic features.

The N4ITK algorithm, widely used to correct for bias field in brain studies, was tested on breast images. As the preset values of the hyperparameters (adapted to cerebral MRI) were not relevant for breast, a grid search of hyperparameters was applied to define corrections relevant for different types of MR breast images. The method was applied to T1-weighted images acquired on a breast phantom and a retrospective dataset of 41 T1-weighted DCE MRI, acquired in our institution with a 1.5 T Magnetom MRI scanner, using two different coils. The impact of the correction was assessed by computing relative image intensities variation between left and right normal tissues in the pectoral muscles (PMr) and the healthy breast parenchyma (BPr), and in the inner and outer layers in the phantom.

New default parameters of N4ITK (5 fitting levels, 50 iterations, mask on the breast region) were necessary for bias field correction. The comparison of PMr and BPr values before and after correction showed a significant reduction after N4ITK (p-value =0.05 for PMr, < 10^-5 for BPr). Similar trends were found on the phantom images.

Retrospective studies need a posteriori algorithms to correct for spatial non-uniformity. Optimised values of N4ITK correction were proposed for breast MRI. Even if perfectible, this fully automatic solution is already useful and should be easily embedded in further radiomic studies.

Further investigations should characterise bias fields for each coil.

No funding was received for this work.

Purpose:

The reconstruction of fibre tracts (FT) of white matter in the brain from diffusion tensor imaging (DTI) data is daily performed at our centre for neurosurgery and stereo-electroencephalography planning in epilepsy patients. Aim of this study is to compare two different algorithms for FT.

We selected 16 patients planned for neurosurgery and acquired with an echo-planar imaging sequence using a receive coil head SENSE 8-channel on a Philips Achieva 1.5T scanner (64 directions, b-value 1000mm²/s). Cortico-spinal tracts were reconstructed with different algorithms using the same input data (i.e. DTI images and regions of interest for seed and waypoints). We used Intellispace Portal v8.0 (Philips Healthcare), a commercial platform for medical image management with a module for the deterministic reconstruction of FT, and the probabilistic algorithm PROBTRACKX by FSL v5.0.6, a free-distributed software for analysis of MRI brain images (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki). For each patient, the tracts were compared in terms of the distance of their centre of gravity (DCoG) and Dice index (DI) calculated within each axial slice of the brain peduncle (BP), internal capsule (IC) and cerebral cortex (CC).

The average DCoG in the BP, IC and CC were, respectively 5.2, 5.5, 9.3mm. The average DI values for the same anatomical sites were 30%, 43%, 12% with values within 0% and 70%, meaning a partial or absence of overlap of the FT reconstructed with the two methods.

The two algorithms for FT shows geometrical differences that cannot be ignored and should be discussed with the neurosurgery team.

Clinical validation of FT reconstructed from DTI acquisition with neurophysiological data is mandatory.

n/a

No funding was received for this work.

Purpose:

To evaluate the clinical applicability and limitations of this new prototype volume-interpolated breath-hold examination (VIBE) with compressed sensing (VIBEcs) for rapid multiphase MRI with a free selectable variable temporal resolution for hypervascularised hepatic lesions.

Twenty patients with hypervascularised hepatic lesions were included in this prospective study and underwent contrast-enhanced liver MRI at 3T. In all patients, VIBEcs was used for rapid arterial multiphase imaging. Results were analysed regarding image quality and clinical applicability of the dynamic lesion evaluation. Evaluation of image quality, visibility and conspicuity were performed by three independent radiologists, each with more than 5 years of experience, based on a 5-point Likert scale (5=excellent). Results were correlated with the lesion entity. Limitations for the use of VIBEcs in image acquisition were defined. Time curves of dynamic contrast enhancement were plotted for each patient and quantification of attenuation performed to isolate the optimal time-point for image acquisition.

All patients were successfully evaluated. Individual setting of acquisition time point (best point 40.6 seconds) instead of fixed delay allowed high reading scores for image quality, visibility and conspicuity for all lesions (mean score >4). Lesion entity showed no significant impact on the reading performance (p=0.765). The limitation was defined as follows: small lesion size (<8mm), subdiaphragmatic localisation, large necrotic area (>80% of the lesions).

Free-breathing MRI with VIBEcs allows image acquisition with high temporal and spatial resolution using individual acquisition time points during contrast phase to gain optimal results with a robust acquisition protocol.

The intention was to test the feasibility. Small patient cohort and a single centre.

The study protocol was approved by the institutional review board. Written obtained consent was obtained by all patients prior to inclusion.

No funding was received for this work.

Purpose:

In white matter, the millimetre-scale resolution of diffusion MRI may encompass myelinated axons and unknown fractions of grey matter, cerebrospinal fluid, or pathological tissue. To tackle this heterogeneity, conventional approaches rely on assumptions regarding tissue properties. This work combines novel MRI acquisition and processing methods to extract voxel-scale nonparametric 5D distributions of diffusion tensors (microstructure) and T₂ (chemical composition) without the use of limiting assumptions. Orientation-resolved (fiber-specific) means of diffusivities and T₂ are then defined via orientation clustering within the 5D distributions.

A healthy volunteer was scanned on a 3T Siemens MAGNETOM Prisma using an EPI sequence customised for tensor-valued diffusion encoding and variable echo times. A nonparametric 5D distribution of diffusion tensors and T₂ is estimated in each voxel via a Monte-Carlo inversion algorithm. This algorithm also provides uncertainty on the estimated distribution and orientation distribution functions (ODFs) of anisotropic components. Finally, orientation-resolved means of isotropic diffusivity ("size"), normalised anisotropy ("shape") and T₂ are computed using an in-house orientation-clustering algorithm.

ODFs present orientations that are consistent with the known anatomy, and are estimated in all areas of the brain, including heterogeneous voxels that comprise white matter, cerebrospinal fluid and/or grey matter. In parallel, our clustering procedure yields the median and interquartile range (uncertainty) of the orientation-resolved means of sizes, shapes and T₂.

We tease apart intra-voxel fibres and separately characterise their respective microstructure and chemical composition. This work shows potential in the understanding and longitudinal tracking of brain development and neurodegenerative diseases.

Pilot study limited by the low number of investigated subjects.

Approved by the IRB of Cardiff University School of Medicine.

Work financially supported by the Swedish Foundation for Strategic Research (ITM17-0267) and the Swedish Research Council (2018-03697).

Purpose:

Conventional diffusion tensor imaging (DTI) metrics are challenging to interpret in terms of specific tissue properties. Here, we address this problem by designing novel acquisition and analysis protocols that quantify the sub-voxel composition of the living brain with nonparametric diffusion tensor distributions (DTDs).

We used a UIH uMR 790 3T MRI system to scan three patients with biopsy-proven pathologies: meningioma, glioblastoma, and intracranial cyst. Data was acquired with a custom diffusion-weighted EPI sequence in less than five minutes and converted to voxel-wise DTDs using a model-free inversion algorithm. The spatially-resolved DTDs were converted to a novel set of 15 parameter maps informing about the diffusivity, anisotropy, and orientation of the microscopic tissue environments.

Healthy tissues are all characterised by distinctive distributions that accurately capture their corresponding diffusion properties; white-matter: low diffusivity, high anisotropy; grey-matter: low diffusivity, low anisotropy; cerebrospinal fluid: high diffusivity, low anisotropy. The investigated pathologies showed different diffusion properties: meningioma presents low diffusivity and high anisotropy, glioblastoma exhibits low diffusivity and low anisotropy, and intracranial cyst displays an isotropic core enveloped by anisotropic tissue. Maps derived from the DTDs allowed a compact visualisation of sub-voxel composition and displayed good agreement with the expected tissue properties.

The proposed framework with spatially-resolved DTDs is capable of resolving and characterising sub-voxel tissue environments that are indistinguishable in conventional DTI. The novel statistical maps derived from the distributions displayed good contrast between different tissues types and provided valuable insight into the structural properties of the studied pathologies.

This pilot study is limited by the low number of investigated patients.

Study approved by the institutional review board of the Wuhan University Hospital.

Work funded by Random Walk Imaging AB and United Imaging Healthcare.

Purpose:

The aim of this study was to evaluate the repeatability and the reproducibility of radiomic features on two 1.5 T MRI scanners of different vendor.

T₂-weighted images of a dedicated pelvis phantom were acquired on scanner A, repeating the acquisitions 10 times without changing the sequence parameters. The phantom was repositioned and the acquisitions repeated. 16 regions of interest were drawn on the phantom inserts. PyRadiomics was used to normalise images and extract radiomic features on original and filtered images. The phantom was acquired on scanner B and the procedure repeated in the same way. To test repeatability - with and without phantom repositioning - the interclass correlation coefficient (ICC) for an absolute agreement was calculated pairwise for each radiomic feature comparing the acquisitions intra-scanner. In order to test reproducibility, the ICC (consistency and absolute agreement) was determined to compare the acquisitions inter-scanner.

Repeatability: the features showing excellent repeatability (ICC > 0.9) were on average 97% for scanner A and 92% for scanner B (no repositioning). The features showing excellent repeatability when repositioning the phantom were 24% (scanner A) and 36% (scanner B). Reproducibility: the features with excellent consistency and agreement were 19% and 10% of the total respectively. The most stable features were those extracted from wavelet filtered images and belonging to the first order and glrlm categories.

The radiomic features showed excellent repeatability when the experiment was performed without any variation. However, less than a quarter of the features proved excellent repeatability when the phantom was repositioned. A selection of the reproducible features should be carried out in case of multicentric studies, given the low reproducibility observed comparing two different scanners.

n/a

n/a

No funding was received for this work.

Purpose:

To obtain a predictive model based on MRI radiomic features for classification of new patients between responders and non-responders to neo-adjuvant chemotherapy (NAC).

60 patients who underwent pre-NAC 3T DCE-MRI and NAC due to breast cancer between January 2015 and October 2018 were included in the study. Two radiologists contoured the lesion in both the pre-contrast and third DCE-MRI scan with a semi-automatic tool. Images were resampled with square voxels (0.9 mm). 214 radiomic features were extracted for each patient, including textural metrics of GLCM, GLRLM, GLSZM, NGTDM and GLDM. Least absolute shrinkage and selection operator (LASSO) was used to select the most suitable grey-level quantization and for feature reduction. Logistic regression and support vector machine (SVM) were used to build a model based on the most significant features. The performance of the multivariable models was assessed by means of ROC analysis, using a leave-one-out cross-validation.

Features including sphericity of the lesion, kurtosis and several higher order metrics turned out to be correlated with the radiological complete response to NAC (p<0.05). The multivariable logistic model, built with the 8 LASSO-selected features, showed sensitivity, specificity, and area under the ROC curve of 0.74, 0.79, and 0.76, respectively, whereas for the SVM model they were 0.67, 0.88, and 0.80.

The presented predictive model provides an indication on the potential of 3T DCE-MRI-based radiomics to predict NAC outcome. Even though the sensitivity of the model is not high enough to grant clinical use, it can provide a substantial support on decision making.

This is a single-institution study.

This study was approved by the institutional review board of the “Azienda Ospedaliera Universitaria Integrata - Verona".

No funding was received for this work.