Preliminary data on a fully automated left ventricular late gadolinium enhancement detection by a convolutional neuronal network in chronic myocardial infarction

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): FWF- Der Wissenschaftsfonds

Aim

To compare a fully automated segmentation of left ventricular late gadolinium enhancement (LGE) as evaluated by a convolutional neuronal network (CNN) with manual segmentation in chronic myocardial infarction.

Methods

Cardiac magnetic resonance imaging including two-dimensional LGE imaging was performed in 191 patients on a 1.5 T clinical scanner 12 months after ST-elevation myocardial infarction. LGE images were presented to a trained CNN for automated determination of left ventricular myocardium and consequently absolute LGE volume. Manual LGE segmentation according to the +5-SD method was used as reference standard. Image quality was assessed according to a 3-point Likert scale (2 = perfect image quality, 1 = some artifacts witout impaired LGE delineation, 0 = strong artifacts with impaired LGE delineation). Regression and Bland-Altman analysis were performed.

Results

In 191 included patients (182 male, mean age 57 years) LGE volume was 9.7 [IQR 3.6 to 16.2] ml according to manual segmentation and 8.3 [3.2 to 17.6] ml according to CNN segmentation. Bland-Altman analysis showed little average difference (-0.5 ml, p=0.257), however, limits of agreement ranged from -18.4 ml to 17.5 ml. Linear correlation was fair (0.57, p<0.001). Subgroup analysis according to image quality showed comparable performance of CNN segmentation in all three groups.

Conclusion

Our fully automated LGE segmentation based on a CNN in two-dimensional data sets provides measurements with little average difference compared to very time-consuming manual segmentations. However, dispersion is substantially and limits the current application of this approach on a per-patient basis. Image quality does not affect CNN performance.

Breaking the resolution limit in photoacoustic imaging using non-negativity and sparsity

The spatial resolution achievable in photoacoustic imaging decreases with the imaging depth, resulting in blurred images for deeper structures. Apart from technical limitations, the ultimate resolution limit results from the second law of thermodynamics. The attenuation of the optically generated acoustic waves on their way from the imaged structure to the sample surface by scattering and dissipation leads to an increase of entropy. The resulting loss of spatial resolution for structures embedded in attenuating media can be compensated by numerical methods that make use of additional available information. In this article, we demonstrate this using experimental data from plane one-dimensional (1D) acoustic waves propagating in fat tissue. The acoustic waves are optically induced by nanosecond laser pulses and measured with piezoelectric transducers. The experimental results of 1D compensation are also relevant for photoacoustic imaging in 2D or 3D in an acoustically attenuating medium by dividing the reconstruction problem into two steps: First, the ideal signal, which is the solution of the un-attenuated wave equation, is determined by the proposed 1D attenuation compensation for each detector signal. In a second step, any ultrasound reconstruction method for un-attenuated data can be used for image reconstruction. For the reconstruction of a small step milled into a silicon wafer surface, which allows the generation of two photoacoustic pulses with a small time offset, we take advantage of non-negativity and sparsity and inverted the measured, frequency dependent acoustic attenuation of the fat tissue. We were able to improve the spatial resolution for imaging through 20 mm of porcine fat tissue compared to the diffraction limit at the cut-off frequency by at least a factor of two.

Photoacoustic reconstruction from photothermal measurements including prior information

Photothermal measurements with an infrared camera enable a fast and contactless part inspection. The main drawback of existing reconstruction methods is the degradation of the spatial resolution with increasing imaging depth, which results in blurred images for deeper lying structures. In this paper, we propose an efficient image reconstruction strategy that allows prior information to be included to overcome the diffusion-based information loss. Following the virtual wave concept, in a first step we reconstruct an acoustic wave field that satisfies the standard wave equation. Therefore, in the second step, stable and efficient reconstruction methods developed for photoacoustic tomography can be used. We compensate for the loss of information in thermal measurements by incorporating the prior information positivity and sparsity. Therefore, we combine circular projections with an iterative regularization scheme. Using simulated and experimental data, this work demonstrates that the quality of the reconstruction from photothermal measurements can be significantly enhanced.

Neural networks-based regularization for large-scale medical image reconstruction

In this paper we present a generalized Deep Learning-based approach for solving ill-posed large-scale inverse problems occuring in medical image reconstruction. Recently, Deep Learning methods using iterative neural networks (NNs) and cascaded NNs have been reported to achieve state-of-the-art results with respect to various quantitative quality measures as PSNR, NRMSE and SSIM across different imaging modalities. However, the fact that these approaches employ the application of the forward and adjoint operators repeatedly in the network architecture requires the network to process the whole images or volumes at once, which for some applications is computationally infeasible. In this work, we follow a different reconstruction strategy by strictly separating the application of the NN, the regularization of the solution and the consistency with the measured data. The regularization is given in the form of an image prior obtained by the output of a previously trained NN which is used in a Tikhonov regularization framework. By doing so, more complex and sophisticated network architectures can be used for the removal of the artefacts or noise than it is usually the case in iterative NNs. Due to the large scale of the considered problems and the resulting computational complexity of the employed networks, the priors are obtained by processing the images or volumes as patches or slices. We evaluated the method for the cases of 3D cone-beam low dose CT and undersampled 2D radial cine MRI and compared it to a total variation-minimization-based reconstruction algorithm as well as to a method with regularization based on learned overcomplete dictionaries. The proposed method outperformed all the reported methods with respect to all chosen quantitative measures and further accelerates the regularization step in the reconstruction by several orders of magnitude.

MatInspector and beyond: promoter analysis based on transcription factor binding sites

MOTIVATION

Promoter analysis is an essential step on the way to identify regulatory networks. A prerequisite for successful promoter analysis is the prediction of potential transcription factor binding sites (TFBS) with reasonable accuracy. The next steps in promoter analysis can be tackled only with reliable predictions, e.g. finding phylogenetically conserved patterns or identifying higher order combinations of sites in promoters of co-regulated genes.

RESULTS

We present a new version of the program MatInspector that identifies TFBS in nucleotide sequences using a large library of weight matrices. By introducing a matrix family concept, optimized thresholds, and comparative analysis, the enhanced program produces concise results avoiding redundant and false-positive matches. We describe a number of programs based on MatInspector allowing in-depth promoter analysis (DiAlignTF, FrameWorker) and targeted design of regulatory sequences (SequenceShaper).

Identification of endogenous retroviral sequences based on modular organization: proviral structure at the SSAV1 locus

The current genome sequencing projects reveal megabases of unknown genomic sequences. About 1% of these sequences can be expected to be of retroviral origin. These are often severely deleted or mutated. Therefore, identification of the retroviral origin of these sequences can be very difficult due to the absence of convincing overall sequence similarity. There are also many copies of solo-LTRs (long terminal repeats) distributed throughout genomic sequences. LTR and envelope sequences in general are among the most divergent parts of the retroviral genome and thus especially hard to detect in mutated endogenous sequences. We took advantage of the fact that these retroviral sections contain short highly conserved sequence regions providing retroviral hallmarks even after loss of overall similarity. We defined several sequence elements and peptide motifs within LTR and Env sequences and used these elements to construct models for LTRs and Env proteins of mammalian C-type retroviruses. We then used this strategy to identify successfully the hitherto missing LTRs and an env-like region in the S71 human retroviral sequence. Our approach provides a new strategy for identifying remotely related retroviral sequences in genomic DNA (especially human DNA), of potential significance for the interpretation of genomic sequences obtained from the current large-scale sequencing projects.

Identification of S71-related human endogenous retroviral sequences with full-length pol genes

The human genome contains sequences related to the simian sarcoma-associated virus SSAV. One of these endogenous retroviral elements, S71, is truncated in the pol gene and carries an insertion of a solitary HERV-K LTR. Using a PCR approach we have now identified further S71-related retroviral elements that lack the HERV-K LTR insertion and contain a full-length retroviral reverse transcriptase. Two of these sequences, pCRTK1 and pCRTK6, were cloned and further characterized. Clones pCRTK1 and pCRTK6 showed between 85 and 90% nucleotide homology to each other and to S71 within the "tether" region of the pol gene, indicating that pCRTK1 and pCRTK6 clearly belong to the S71 subgroup of C-type-related human endogenous retroviral elements. Some point mutations inactivating the reverse transcriptase are located at the same positions in pCRTK1 and pCRTK6. Therefore, we assume that these S71-related elements were dispersed in the human genome by reintegration as defective proviruses, probably using enzymes for retrotransposition provided in trans by other retrotransposons or by cellular genes. Examination of the presence of S71-related elements in apes and Old World monkeys revealed that the deletion of reverse transcriptase sequences in S71 has occurred in the lineage of primates prior to the insertion of the HERV-K LTR.

Transcription of HERV-K-related LTRs in human placenta and leukemic cells

The human genome contains a family of endogenous retroviruses, HERV-K, with sequence homology to the B-type mouse mammary tumor virus (MMTV). We have detected HERV-K-LTR related cDNA clones by screening a human placenta cDNA library with a HERV-K LTR probe. Three of the isolated cDNA clones were characterized by nucleotide sequencing. The analyzed clones did not contain any retroviral sequences other than those related to HERV-K LTRs, but were found to be coexpressed with cellular sequences. Furthermore, transcripts containing HERV-K LTR sequences were demonstrated by Northern blotting and PCR in human leukemic and normal white blood cells, as well as in various tumor cell lines, indicating abundant transcription of solitary HERV-K LTRs in human tissues. In patients with lymphatic leukemias, a transcript of about 6 kb hybridizing with HERV-K LTR was detected that was not found in patients with myelogenous leukemias or in healthy persons.

Genomic distribution and transcription of solitary HERV-K LTRs

The human genome contains a family of endogenous retroviruses, HERV-K, with sequence homology to the B-type mouse mammary tumor virus. We have now identified a single HERV-K LTR within the C-type-related human retroviral element S71. The HERV-K LTR is located in the antisense direction between the S71 gag and the pol gene, replacing the 5' half of S71 pol. A number of HERV-K LTR-related cDNA clones were detected by screening various human cDNA libraries with an S71 HERV-K LTR probe, indicating abundant transcription of HERV-K-related LTRs in human tissues. Sequence analysis of four cDNA clones revealed LTR sequences with a nucleotide identity of 70 to 90% with HERV-K10 LTR. Some HERV-K-related LTR sequences contain potential short open reading frames. The analyzed cDNA clones do not harbor any retroviral sequences other than those related to HERV-K LTRs. However, most of the solitary LTRs were found to be coexpressed with cellular sequences. Transcription of these LTRs is probably directed by external cellular promoters. We show that HERV-KLTR-like sequences entered the primate genome about 33-40 million years ago. We estimate the human genome to contain about 25,000 copies of HERV-K-related LTRs, which are distributed over most human chromosomes in an irregular manner.