Unsupervised lesion detection via image restoration with a normative prior.

Unsupervised lesion detection is a challenging problem that requires accurately estimating normative distributions of healthy anatomy and detecting lesions as outliers without training examples. Recently, this problem has received increased attention from the research community following the advances in unsupervised learning with deep learning. Such advances allow the estimation of high-dimensional distributions, such as normative distributions, with higher accuracy than previous methods. The main approach of the recently proposed methods is to learn a latent-variable model parameterized with networks to approximate the normative distribution using example images showing healthy anatomy, perform prior-projection, i.e. reconstruct the image with lesions using the latent-variable model, and determine lesions based on the differences between the reconstructed and original images. While being promising, the prior-projection step often leads to a large number of false positives. In this work, we approach unsupervised lesion detection as an image restoration problem and propose a probabilistic model that uses a network-based prior as the normative distribution and detect lesions pixel-wise using MAP estimation. The probabilistic model punishes large deviations between restored and original images, reducing false positives in pixel-wise detections. Experiments with gliomas and stroke lesions in brain MRI using publicly available datasets show that the proposed approach outperforms the state-of-the-art unsupervised methods by a substantial margin, +0.13 (AUC), for both glioma and stroke detection. Extensive model analysis confirms the effectiveness of MAP-based image restoration.

Analysis of spatiotemporal fidelity in quantitative 3D first-pass perfusion cardiovascular magnetic resonance.

BACKGROUND: Whole-heart first-pass perfusion cardiovascular magnetic resonance (CMR) relies on highly accelerated image acquisition. The influence of undersampling on myocardial blood flow (MBF) quantification has not been systematically investigated yet. In the present work, the effect of spatiotemporal scan acceleration on image reconstruction accuracy and MBF error was studied using a numerical phantom and validated in-vivo. METHODS: Up to 10-fold scan acceleration using k-t PCA and k-t SPARSE-SENSE was simulated using the MRXCAT CMR numerical phantom framework. Image reconstruction results were compared to ground truth data in the k-f domain by means of modulation transfer function (MTF) analysis. In the x-t domain, errors pertaining to specific features of signal intensity-time curves and MBF values derived using Fermi model deconvolution were analysed. In-vivo first-pass CMR data were acquired in ten healthy volunteers using a dual-sequence approach assessing the arterial input function (AIF) and myocardial enhancement. 10x accelerated 3D k-t PCA and k-t SPARSE-SENSE were compared and related to non-accelerated 2D reference images. RESULTS: MTF analysis revealed good recovery of data upon k-t PCA reconstruction at 10x undersampling with some attenuation of higher temporal frequencies. For 10x k-t SPARSE-SENSE the MTF was found to decrease to zero at high spatial frequencies for all temporal frequencies indicating a loss in spatial resolution. Signal intensity-time curve errors were most prominent in AIFs from 10x k-t PCA, thereby emphasizing the need for separate AIF acquisition using a dual-sequence approach. These findings were confirmed by MBF estimation based on AIFs from fully sampled and undersampled simulations. Average in-vivo MBF estimates were in good agreement between both accelerated and the fully sampled methods. Intra-volunteer MBF variation for fully sampled 2D scans was lower compared to 10x k-t PCA and k-t SPARSE-SENSE data. CONCLUSION: Quantification of highly undersampled 3D first-pass perfusion CMR yields accurate MBF estimates provided the AIF is obtained using fully sampled or moderately undersampled scans as part of a dual-sequence approach. However, relative to fully sampled 2D perfusion imaging, intra-volunteer variation is increased using 3D approaches prompting for further developments.

Single-Molecule Localization Microscopy allows for the analysis of cancer metastasis-specific miRNA distribution on the nanoscale.

We describe a novel approach for the detection of small non-coding RNAs in single cells by Single-Molecule Localization Microscopy (SMLM). We used a modified SMLM-setup and applied this instrument in a first proof-of-principle concept to human cancer cell lines. Our method is able to visualize single microRNA (miR)-molecules in fixed cells with a localization accuracy of 10-15 nm, and is able to quantify and analyse clustering and localization in particular subcellular sites, including exosomes. We compared the metastasis-site derived (SW620) and primary site derived (SW480) human colorectal cancer (CRC) cell lines, and (as a proof of principle) evaluated the metastasis relevant miR-31 as a first example. We observed that the subcellular distribution of miR-31 molecules in both cell lines was very heterogeneous with the largest subpopulation of optically acquired weakly metastatic cells characterized by a low number of miR-31 molecules, as opposed to a significantly higher number in the majority of the highly metastatic cells. Furthermore, the highly metastatic cells had significantly more miR-31-molecules in the extracellular space, which were visualized to co-localize with exosomes in significantly higher numbers. From this study, we conclude that miRs are not only aberrantly expressed and regulated, but also differentially compartmentalized in cells with different metastatic potential. Taken together, this novel approach, by providing single molecule images of miRNAs in cellulo can be used as a powerful supplementary tool in the analysis of miRNA function and behaviour and has far reaching potential in defining metastasis-critical subpopulations within a given heterogeneous cancer cell population.