Relational local electroencephalography representations for sleep scoring.

Computational sleep scoring from multimodal neurophysiological time-series (polysomnography PSG) has achieved impressive clinical success. Models that use only a single electroencephalographic (EEG) channel from PSG have not yet received the same clinical recognition, since they lack Rapid Eye Movement (REM) scoring quality. The question whether this lack can be remedied at all remains an important one. We conjecture that predominant Long Short-Term Memory (LSTM) models do not adequately represent distant REM EEG segments (termed epochs), since LSTMs compress these to a fixed-size vector from separate past and future sequences. To this end, we introduce the EEG representation model ENGELBERT (electroEncephaloGraphic Epoch Local Bidirectional Encoder Representations from Transformer). It jointly attends to multiple EEG epochs from both past and future. Compared to typical token sequences in language, for which attention models have originally been conceived, overnight EEG sequences easily span more than 1000 30 s epochs. Local attention on overlapping windows reduces the critical quadratic computational complexity to linear, enabling versatile sub-one-hour to all-day scoring. ENGELBERT is at least one order of magnitude smaller than established LSTM models and is easy to train from scratch in a single phase. It surpassed state-of-the-art macro F1-scores in 3 single-EEG sleep scoring experiments. REM F1-scores were pushed to at least 86%. ENGELBERT virtually closed the gap to PSG-based methods from 4-5 percentage points (pp) to less than 1 pp F1-score.

PET image segmentation using a Gaussian mixture model and Markov random fields.

BACKGROUND: Classification algorithms for positron emission tomography (PET) images support computational treatment planning in radiotherapy. Common clinical practice is based on manual delineation and fixed or iterative threshold methods, the latter of which requires regression curves dependent on many parameters. METHODS: An improved statistical approach using a Gaussian mixture model (GMM) is proposed to obtain initial estimates of a target volume, followed by a correction step based on a Markov random field (MRF) and a Gibbs distribution to account for dependencies among neighboring voxels. In order to evaluate the proposed algorithm, phantom measurements of spherical and non-spherical objects with the smallest diameter being 8 mm were performed at signal-to-background ratios (SBRs) between 2.06 and 9.39. Additionally (68)Ga-PET data from patients with lesions in the liver and lymph nodes were evaluated. RESULTS: The proposed algorithm produces stable results for different reconstruction algorithms and different lesion shapes. Furthermore, it outperforms all threshold methods regarding detection rate, determines the spheres' volumes more accurately than fixed threshold methods, and produces similar values as iterative thresholding. In a comparison with other statistical approaches, the algorithm performs equally well for larger volumes and even shows improvements for small volumes and SBRs. The comparison with experts' manual delineations on the clinical data shows the same qualitative behavior as for the phantom measurements. CONCLUSIONS: In conclusion, a generic probabilistic approach that does not require data measured beforehand is presented whose performance, robustness, and swiftness make it a feasible choice for PET segmentation.

Automatic ictal HFO detection for determination of initial seizure spread.

High-frequency oscillations (HFOs) are a reliable indicator for the epileptic seizure onset zone (SOZ) in ECoG recordings. We propose a novel method for the automatic detection of ictal HFOs in the ripple band (80-250 Hz) based on CFAR matched sub-space filtering. This allows to track the early propagation of ictal HFOs, revealing initial and follow-up epileptic activity on the electrodes. We apply this methodology to two seizures from one patient suffering from focal epilepsy. The electrodes identified are in very good accordance with the visual HFO analysis by clinicians. Furthermore the electrodes with initial HFO activity are correlated well with the SOZ (conventional v-activity).

Artefact reduction for cell migration visualization using spectral domain optical coherence tomography.

Visualization of cell migration during chemotaxis using spectral domain optical coherence tomography (OCT) requires non-standard processing techniques. Stripe artefacts and camera noise floor present in OCT data prevent detailed computer-assisted reconstruction and quantification of cell locomotion. Furthermore, imaging artefacts lead to unreliable results in automated texture based cell analysis. Here we characterize three pronounced artefacts that become visible when imaging sample structures with high dynamic range, e.g. cultured cells: (i) time-varying fixed-pattern noise; (ii) stripe artefacts generated by background estimation using tomogram averaging; (iii) image modulations due to spectral shaping. We evaluate techniques to minimize the above mentioned artefacts using an 800 nm optical coherence microscope. Effect of artefact reduction is shown exemplarily on two cell cultures, i.e. Dictyostelium on nitrocellulose substrate, and retinal ganglion cells (RGC-5) cultured on a glass coverslip. Retinal imaging also profits from the proposed processing techniques.

Fast dispersion encoded full range optical coherence tomography for retinal imaging at 800 nm and 1060 nm.

The dispersion mismatch between sample and reference arm in frequency-domain optical coherence tomography (OCT) can be used to iteratively suppress complex conjugate artifacts and thereby increase the imaging range. In this paper, we propose a fast dispersion encoded full range (DEFR) algorithm that detects multiple signal components per iteration. The influence of different dispersion levels on the reconstruction quality is analyzed experimentally using a multilayered scattering phantom and in vivo retinal tomograms at 800 nm. Best results have been achieved with 30 mm SF11, with neglectable resolution decrease due to finite resolution of the spectrometer. Our fast DEFR algorithm achieves an average suppression ratio of 55 dB and typically converges within 5 to 10 iterations. The processing time on non-dedicated hardware was 5 to 10 seconds for tomograms with 512 depth scans and 4096 sampling points per depth scan. Application of DEFR to the more challenging 1060 nm wavelength region is also demonstrated by introducing an additional optical fibre in the sample arm.

Dispersion encoded full range frequency domain optical coherence tomography.

We propose an iterative algorithm that exploits the dispersion mismatch between reference and sample arm in frequency-domain optical coherence tomography (FD-OCT) to effectively cancel complex conjugate mirror terms in individual A-scans and thereby generate full range tomograms. The resulting scheme, termed dispersion encoded full range (DEFR) OCT, allows distinguishing real structures from complex conjugate mirror artifacts. Even though DEFR-OCT has higher post-processing complexity than conventional FD-OCT, acquisition speed is not compromised since no additional A-scans need to be measured, thereby rendering this technique robust against phase fluctuations. The algorithm uses numerical dispersion compensation and exhibits similar resolution as standard processing. The residual leakage of mirror terms is further reduced by incorporating additional knowledge such as the power spectrum of the light source. The suppression ratio of mirror signals is more than 50 dB and thus comparable to complex FD-OCT techniques which use multiple A-scans.