Machine learning methods offer great promise for fast and accurate detection and prognostication of coronavirus disease 2019 (COVID-19) from standard-of-care chest radiographs (CXR) and chest computed tomography (CT) images. Many articles have been published in 2020 describing new machine learning-based models for both of these tasks, but it is unclear which are of potential clinical utility. In this systematic review, we consider all published papers and preprints, for the period from 1 January 2020 to 3 October 2020, which describe new machine learning models for the diagnosis or prognosis of COVID-19 from CXR or CT images. All manuscripts uploaded to bioRxiv, medRxiv and arXiv along with all entries in EMBASE and MEDLINE in this timeframe are considered. Our search identified 2,212 studies, of which 415 were included after initial screening and, after quality screening, 62 studies were included in this systematic review. Our review finds that none of the models identified are of potential clinical use due to methodological flaws and/or underlying biases. This is a major weakness, given the urgency with which validated COVID-19 models are needed. To address this, we give many recommendations which, if followed, will solve these issues and lead to higher-quality model development and well-documented manuscripts.

Förster resonance energy transfer (FRET) detected via fluorescence lifetime imaging microscopy (FLIM) and global analysis provide a way in which protein-protein interactions may be spatially localized and quantified within biological cells. The FRET efficiency and proportion of interacting molecules have been determined using bi-exponential fitting to time-domain FLIM data from a multiphoton time-correlated single-photon counting microscope system. The analysis has been made more robust to noise and significantly faster using global fitting, allowing higher spatial resolutions and/or lower acquisition times. Data have been simulated, as well as acquired from cell experiments, and the accuracy of a modified Levenberg-Marquardt fitting technique has been explored. Multi-image global analysis has been used to follow the epidermal growth factor-induced activation of Cdc42 in a short-image-interval time-lapse FLIM/FRET experiment. Our implementation offers practical analysis and time-resolved-image manipulation, which have been targeted towards providing fast execution, robustness to low photon counts, quantitative results and amenability to automation and batch processing.

FLIM/FRET is an extremely powerful technique that can microscopically locate nanometre-scale protein-protein interactions within live or fixed cells, both in vitro and in vivo. The key to performing sensitive FRET, via FLIM, besides the use of appropriate fluorophores, is the analysis of the time-resolved data present at each image pixel. The fluorescent transient will, in general, exhibit multi-exponential kinetics: at least two exponential components arise from both the interacting and non-interacting protein. We shall describe a novel method and computer program for the global analysis of time resolved data, either at the single level or through global analysis of grouped pixel data. Kinetic models are fitted using the Marquardt algorithm and iterative convolution of the excitation signal, in a computationally-efficient manner. The fitting accuracy and sensitivity of the algorithm has been tested using modelled data, including the addition of simulated Poisson noise and repetitive excitation pulses which are typical of a TCSPC system. We found that the increased signal to noise ratio offered by both global and invariance fitting is highly desirable. When fitting mono-exponential data, the effects of a ca.12.5 ns (ca. 80 MHz) repetitive excitation do not preclude the accurate extraction of populations with lifetimes in the range 0.1 to 10 ns, even when these effects are not represented in the fitting algorithm. Indeed, with global or invariance fitting of a 32x32 pixel area, the error in extracted lifetime can be lower than 0.4 % for signals with a peak of 500 photon counts or more. In FRET simulations, modelling GFP with a non-interacting lifetime of 2.15 ns, it was possible to accurately detect a 10 % interacting population with a lifetime of 0.8 ns.

The application of a cost-effective spectral imager to spatially segmenting absorptive and fluorescent chemical probes on the basis of their spectral characteristics has been demonstrated. The imager comprises a computer-controlled spectrally selective element that allows random access to a bandwidth of 15 nm between 400 and 700 nm. Further, the use of linear un-mixing of the spectral response of a sample at a single pixel has been facilitated using non-negative least squares fitting. Examples are given showing the separation of dye distributions, such as immunohistochemical markers for tumour hypoxia, from multiply stained thin tissue sections, imaged by trans-illumination microscopy. A quantitative study is also presented that shows a correlation between staining intensity and normal versus tumour tissue, and the advantage of reducing the amount of data captured for a particular study is also demonstrated. An example of the application to fluorescence microscopy is also given, showing the separation of green fluorescent protein, Cy3 and Cy5 at a single pixel. The system has been validated against samples of known optical density and of known overlapping combinations of coloured filters. These results confirm the ability of this technique to separate spectral responses that cannot be resolved with conventional colour imaging.

No abstract

No abstract

Classifying samples in incomplete datasets is a common aim for machine learning practitioners, but is non-trivial. Missing data is found in most real-world datasets and these missing values are typically imputed using established methods, followed by classification of the now complete, imputed, samples. The focus of the machine learning researcher is then to optimise the downstream classification performance. In this study, we highlight that it is imperative to consider the quality of the imputation. We demonstrate how the commonly used measures for assessing quality are flawed and propose a new class of discrepancy scores which focus on how well the method recreates the overall distribution of the data. To conclude, we highlight the compromised interpretability of classifier models trained using poorly imputed data. All code and data used in this paper are also released publicly at [inserted upon publication].

Abstract. We consider the permutation group algebra defined by Cameron and show that if the permutation group has no finite orbits, then no homogeneous element of degree one is a zero-divisor of the algebra. We proceed to make a conjecture which would show that the algebra is an integral domain if, in addition, the group is oligomorphic. We go on to show that this conjecture is true in certain special cases, including those of the form H Wr S and H Wr A, and show that in the oligormorphic case, the algebras corresponding to these special groups are polynomial algebras. In the H Wr A case, the algebra is related to the shuffle algebra of free Lie algebra theory.