Search

Background: In patients with soft-tissue sarcomas, tumor grading constitutes a decisive factor to determine the best treatment decision. Tumor grading is obtained by pathological work-up after focal biopsies. Deep learning (DL)-based imaging analysis may pose an alternative way to characterize STS tissue. In this work, we sought to non-invasively differentiate tumor grading into low-grade (G1) and high-grade (G2/G3) STS using DL techniques based on MR-imaging. Methods: Contrast-enhanced T1-weighted fat-saturated (T1FSGd) MRI sequences and fat-saturated T2-weighted (T2FS) sequences were collected from two independent retrospective cohorts (training: 148 patients, testing: 158 patients). Tumor grading was determined following the French Federation of Cancer Centers Sarcoma Group in pre-therapeutic biopsies. DL models were developed using transfer learning based on the DenseNet 161 architecture. Results: The T1FSGd and T2FS-based DL models achieved area under the receiver operator characteristic curve (AUC) values of 0.75 and 0.76 on the test cohort, respectively. T1FSGd achieved the best F1-score of all models (0.90). The T2FS-based DL model was able to significantly risk-stratify for overall survival. Attention maps revealed relevant features within the tumor volume and in border regions. Conclusions: MRI-based DL models are capable of predicting tumor grading with good reproducibility in external validation.

The purpose of this study was to analyze size and growth dynamics of focal lesions (FL) as well as to quantify diffuse infiltration (DI) in untreated smoldering multiple myeloma (SMM) patients and correlate those MRI features with timepoint and cause of progression. We investigated 199 whole-body magnetic resonance imaging (wb-MRI) scans originating from longitudinal imaging of 60 SMM patients and 39 computed tomography (CT) scans for corresponding osteolytic lesions (OL) in 17 patients. All FLs >5 mm were manually segmented to quantify volume and growth dynamics, and DI was scored, rating four compartments separately in T1- and fat-saturated T2-weighted images. The majority of patients with at least two FLs showed substantial spatial heterogeneity in growth dynamics. The volume of the largest FL (p = 0.001, c-index 0.72), the speed of growth of the fastest growing FL (p = 0.003, c-index 0.75), the DI score (DIS, p = 0.014, c-index 0.67), and its dynamic over time (DIS dynamic, p < 0.001, c-index 0.67) all significantly correlated with the time to progression. Size and growth dynamics of FLs correlated significantly with presence/appearance of OL in CT within 2 years after the respective MRI assessment (p = 0.016 and p = 0.022). DIS correlated with decrease of hemoglobin (p < 0.001). In conclusion, size and growth dynamics of FLs correlate with prognosis and local bone destruction. Connections between MRI findings and progression patterns (fast growing FL—OL; DIS—hemoglobin decrease) might enable more precise diagnostic and therapeutic approaches for SMM patients in the future.

This thesis takes MRI a step further to study brain tissue microstructure. The dMRI signal is reformulated in a Blind Source Separation framework, enabling the disentanglement of sub-voxel tissue signal components, and the estimation of multiple tissue parameters. Furthermore, a deep learning model is introduced, tackling the partial volume contamination caused by Cerebrospinal Fluid in dMRI. Finally, Quantitative Transient-state Imaging, an ultra-fast acquisition and reconstruction scheme for multiparameter mapping, is extended to a tissue multicompartment model., Diese Arbeit führt die MRT einen Schritt weiter, um die Mikrostruktur des Hirngewebes zu untersuchen. Das dMRI-Signal wird in einem Blind-Source-Separation-Framework umformuliert, das die Entflechtung von Sub-Voxel-Gewebe-Signalkomponenten und die Schätzung mehrerer Gewebeparameter ermöglicht. Darüber hinaus wird ein Deep-Learning-Modell eingeführt, das die partielle Volumenkontamination von Liquor in dMRI in Angriff nimmt. Schließlich wird die quantitative transitorische Bildgebung, ein ultraschnelles Akquisitions- und Rekonstruktionsschema für Multiparameter-Mapping, auf ein Gewebe-Multikompartiment-Modell erweitert.

Novel methods for quantitative, transient-state multiparametric imaging are increasingly being demonstrated for assessment of disease and treatment efficacy. Here, we build on these by assessing the most common Non-Cartesian readout trajectories (2D/3D radials and spirals), demonstrating efficient anti-aliasing with a k-space view-sharing technique, and proposing novel methods for parameter inference with neural networks that incorporate the estimation of proton density. Our results show good agreement with gold standard and phantom references for all readout trajectories at 1.5 T and 3 T. Parameters inferred with the neural network were within 6.58% difference from the parameters inferred with a high-resolution dictionary. Concordance correlation coefficients were above 0.92 and the normalized root mean squared error ranged between 4.2 and 12.7% with respect to gold-standard phantom references for T1 and T2. In vivo acquisitions demonstrate sub-millimetric isotropic resolution in under five minutes with reconstruction and inference times < 7 min. Our 3D quantitative transient-state imaging approach could enable high-resolution multiparametric tissue quantification within clinically acceptable acquisition and reconstruction times.

Whole-body imaging of mice is a key source of information for research. Organ segmentation is a prerequisite for quantitative analysis but is a tedious and error-prone task if done manually. Here, we present a deep learning solution called AIMOS that automatically segments major organs (brain, lungs, heart, liver, kidneys, spleen, bladder, stomach, intestine) and the skeleton in less than a second, orders of magnitude faster than prior algorithms. AIMOS matches or exceeds the segmentation quality of state-of-the-art approaches and of human experts. We exemplify direct applicability for biomedical research for localizing cancer metastases. Furthermore, we show that expert annotations are subject to human error and bias. As a consequence, we show that at least two independently created annotations are needed to assess model performance. Importantly, AIMOS addresses the issue of human bias by identifying the regions where humans are most likely to disagree, and thereby localizes and quantifies this uncertainty for improved downstream analysis. In summary, AIMOS is a powerful open-source tool to increase scalability, reduce bias, and foster reproducibility in many areas of biomedical research.

This study aims to develop a silent, fast and 3D method for T1 and proton density (PD) mapping, while generating time series of T1-weighted (T1w) images with bias-field correction. Undersampled T1w images at different effective inversion times (TIs) were acquired using the inversion recovery prepared RUFIS sequence with an interleaved k-space trajectory. Unaliased images were reconstructed by constraining the signal evolution to a temporal subspace which was learned from the signal model. Parameter maps were obtained by fitting the data to the signal model, and bias-field correction was conducted on T1w images. Accuracy and repeatability of the method was accessed in repeated experiments with phantom and volunteers. For the phantom study, T1 values obtained by the proposed method were highly consistent with values from the gold standard method, R2 = 0.9976. Coefficients of variation (CVs) ranged from 0.09% to 0.83%. For the volunteer study, T1 values from gray and white matter regions were consistent with literature values, and peaks of gray and white matter can be clearly delineated on whole-brain T1 histograms. CVs ranged from 0.01% to 2.30%. The acoustic noise measured at the scanner isocenter was 2.6 dBA higher compared to the in-bore background. Rapid and with low acoustic noise, the proposed method is shown to produce accurate T1 and PD maps with high repeatability by reconstructing sparsely sampled T1w images at different TIs using temporal subspace. Our approach can greatly enhance patient comfort during examination and therefore increase the acceptance of the procedure.

The purpose of this study was to assess how different MRI protocols (spinal vs. spinal plus pelvic vs. whole-body (wb)-MRI) affect staging in patients with smoldering multiple myeloma (SMM), according to the SLiM-CRAB-criterion ‘>1 focal lesion (FL) in MRI’. In this retrospective study, a baseline cohort of 147 SMM patients with wb-MRI at initial diagnosis was investigated, including prognostic data regarding development of CRAB-criteria. Fifty-two patients formed a follow-up cohort with a median of three wb-MRIs. The locations of all FLs were determined and it was calculated how staging decisions regarding the criterion ‘>1 FL in MRI’ would have been made if only a limited anatomic area (spine vs. spine plus pelvis) would have been covered by the MRI protocol. Furthermore, subgroups of patients selected by different cutoff-protocol-combinations were compared regarding their prognosis for development of CRAB-criteria. With an MRI protocol limited to spine/spine plus pelvis, only 28%/64% of patients who actually had >1 FL in wb-MRI would have been rated correctly as having ‘>1 FL in MRI’. Fifty-four percent/36% of patients with exactly 1 FL in spine/spine plus pelvis revealed >1 FL when the entire wb-MRI was analyzed. During follow-up, four more patients developed >1 FL in wb-MRI; both limited MRI protocols would have detected only one of these four patients as having >1 FL at the correct timepoint. Having >1 FL in spine/in spine plus pelvis/in the whole body was associated with a 43%/57%/49% probability of developing CRAB-criteria within 2 years. Patients with >3 FL in spine plus pelvis and patients with >4 FL in the whole body had an 80% probability to develop CRAB-criteria within 2 years. MRI protocols limited to the spine or to spine plus pelvis lead to substantial underdiagnoses of patients who actually have >1 FL in wb-MRI at baseline and during follow-up, which influences staging and treatment decisions according to the current

This work focuses on developing novel methods to accelerate Magnetic Resonance Imaging (MRI), to overcome its persistent fundamental challenge: the acquisition time. To this end this thesis presents scientific contributions in three areas of quantitative MRI: hyperpolarized 13C metabolic imaging in preclinical rodent models of cancer, motion encoding methods to map diffusion and flow in vivo, and multiparametric mapping techniques for fast, quantitative neuroimaging in the human brain., Die vorliegende Arbeit konzentriert sich auf die Entwicklung neuer Methoden zur Beschleunigung der Magnetresonanztomographie (MRT) und adressiert damit eine ihrer grundsätzlichen Herausforderungen: die Messzeit. Dazu werden wissenschaftliche Beiträge in drei Bereichen der quantitativen MRT präsentiert: hyperpolarisierte 13C metabolische Bildgebung, Methoden zur Kodierung von Bewegung, um Diffusion und Blutfluss zu messen sowie multiparametrische Quantifizierungstechniken für die Neurobildgebung.

Correction to: Nature Communications; https://doi.org/10.1038/s41467-018-07619-7; published online 06 December 2018In the original version of this Article the values in the rightmost column of Table 1 were inadvertently shifted relative to the other columns. This has now been corrected in the PDF and HTML versions of the Article.

We propose an auto-encoding network architecture for point clouds (PC) capable of extracting shape signatures without supervision. Building on this, we (i) design a loss function capable of modelling data variance on PCs which are unstructured, and (ii) regularise the latent space as in a variational auto-encoder, both of which increase the auto-encoders' descriptive capacity while making them probabilistic. Evaluating the reconstruction quality of our architectures, we employ them for detecting vertebral fractures without any supervision. By learning to efficiently reconstruct only healthy vertebrae, fractures are detected as anomalous reconstructions. Evaluating on a dataset containing $\sim$1500 vertebrae, we achieve area-under-ROC curve of $>$75%, without using intensity-based features., Comment: Accepted at Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2019; JSK and BHM are joint supervising authors