Image Quality Assessment to Emulate Experts’ Perception in Lumbar MRI Using Machine Learning

Medical image quality is crucial to obtaining reliable diagnostics. Most quality controls rely on routine tests using phantoms, which do not reflect closely the reality of images obtained on patients and do not reflect directly the quality perceived by radiologists. The purpose of this work is to develop a method that classifies the image quality perceived by radiologists in MR images. The focus was set on lumbar images as they are widely used with different challenges. Three neuroradiologists evaluated the image quality of a dataset that included T1-weighting images in axial and sagittal orientation, and sagittal T2-weighting. In parallel, we introduced the computational assessment using a wide range of features extracted from the images, then fed them into a classifier system. A total of 95 exams were used, from our local hospital and a public database, and part of the images was manipulated to broaden the distribution quality of the dataset. Good recall of 82% and an area under curve (AUC) of 77% were obtained on average in testing condition, using a Support Vector Machine. Even though the actual implementation still relies on user interaction to extract features, the results are promising with respect to a potential implementation for monitoring image quality online with the acquisition process.

SARS-CoV-2 infection and acute ischemic stroke in Lombardy, Italy

Objective To characterize patients with acute ischemic stroke related to SARS-CoV-2 infection and assess the classification performance of clinical and laboratory parameters in predicting in-hospital outcome of these patients. Methods In the setting of the STROKOVID study including patients with acute ischemic stroke consecutively admitted to the ten hub hospitals in Lombardy, Italy, between March 8 and April 30, 2020, we compared clinical features of patients with confirmed infection and non-infected patients by logistic regression models and survival analysis. Then, we trained and tested a random forest (RF) binary classifier for the prediction of in-hospital death among patients with COVID-19. Results Among 1013 patients, 160 (15.8%) had SARS-CoV-2 infection. Male sex (OR 1.53; 95% CI 1.06–2.27) and atrial fibrillation (OR 1.60; 95% CI 1.05–2.43) were independently associated with COVID-19 status. Patients with COVID-19 had increased stroke severity at admission [median NIHSS score, 9 (25th to75th percentile, 13) vs 6 (25th to75th percentile, 9)] and increased risk of in-hospital death (38.1% deaths vs 7.2%; HR 3.30; 95% CI 2.17–5.02). The RF model based on six clinical and laboratory parameters exhibited high cross-validated classification accuracy (0.86) and precision (0.87), good recall (0.72) and F1-score (0.79) in predicting in-hospital death. Conclusions Ischemic strokes in COVID-19 patients have distinctive risk factor profile and etiology, increased clinical severity and higher in-hospital mortality rate compared to non-COVID-19 patients. A simple model based on clinical and routine laboratory parameters may be useful in identifying ischemic stroke patients with SARS-CoV-2 infection who are unlikely to survive the acute phase.

Remembering rhythm and rhyme: memorability of narratives for science communication

Once upon a time, narratives were considered to be a non-reliable way of representing and communicating science. Nowadays, narratives are widely accepted as an accurate way of conveying science; they represent an effective emotional trigger, a lasting memory structure and a powerful aid for learning. To study how memorable different ways of presenting information can be is a fundamental task for science communication in order to evaluate materials that not only need to be understood by the general public but also need to be retained in the long term as a part of the communication process. In this paper, I will give a brief introduction to cognitive psychology and the study of memory in relation to narratives. Evidence from the field of memory studies suggests that narratives represent a good recall device. They can generate emotion, and this in turn is a way of focusing attention, promoting rehearsal in memory and inducing long-term potentiation. Similarly, a story produces semantic links that might assist in storing and retrieving information from memory. Studies suggest that memory span and paired-associate recall have implications in storing and recalling narratives. Evidence also suggests that the use of stories as modelling tools can organise information, provide schemas and allow extrapolation or prediction. Finally, literature in memory suggests that narratives have value as mnemonic devices.

Adolescents with congenital heart defects: a patient and parental perspective of genetic information and genetic risk

Congenital heart defects (CHDs) occur in 8 of 1000 live-born children, making them common birth defects in the adolescent population. CHDs may have single gene, chromosomal, or multifactorial causes. Despite evidence that patients with CHD want information on heritability and genetics, no studies have investigated the interest or knowledge base in the adolescent population. This information is necessary as patients in adolescence take greater ownership of their health care and discuss reproductive risks with their physicians. The objectives of this survey-based study were to determine adolescents’ recall of their own heart condition, to assess patient and parent perception of the genetic contribution to the adolescent’s CHD, and to obtain information about the preferred method(s) for education. The results show that adolescent patients had good recall of their type of CHD. Less than half of adolescents and parents believed their CHD had a genetic basis or was heritable; however, adolescents with a positive family history of CHD were more likely to believe that their condition was genetic (p = 0.0005). The majority of patients were interested in receiving additional genetics education and preferred education in-person and in consultation with both parents and a physician. The adolescents who felt most competent to have discussions with their doctors regarding potential causes of their heart defect previously had a school science course which covered topics in genetics. These results provide insight into adolescents’ perceptions and understanding about their CHD and genetic risk and may inform the creation and provision of additional genetic education.

Intelli-NGS: Intelligent NGS, a deep neural network-based artificial intelligence to delineate good and bad variant calls from IonTorrent sequencer data

BackgroundIonTorrent is a second-generation sequencing platform with smaller capital costs than Illumina but is also prone to higher machine error than later. Given its lower costs, the platform is generally preferred in developing countries where next-generation sequencing is still a very exclusive technique. There are many software tools available for other platforms but IonTorrent. This makes the already tricky analysis part more error-prone.MotivationWe have been using the IonTorrent platform in our hospital setting for aiding diagnosis or treatment for the past couple of years. Given to our experience, analysis part of IonTorrent data takes the longest time and still, we used to get stuck with certain variants which seemed fine on looking at their metrics but were found to be negative in Sanger sequencing verification. This made us determined to develop a tool that could aid us in reducing false positive and negative rates while still retaining good recall. The artificial intelligence-based technique was our final choice after developing pipelines with less success.MethodologyThe artificial intelligence was developed from scratch in Python 3 using TensorFlow fully connected dense layers. The model takes VCF files as input and solves each variant based on the thirty-five parameters given by the IonTorrent platform, including the flow-space information which is missed by variant callers other than the default torrent variant caller.ResultsThe final trained model was able to achieve an accuracy of 93.08% and a ROC-AUC of 0.95 with GIAB validation data. The additional program that was written to run the model annotates each variant using online databases such as dbSNP, ClinVar and others. A probability score for each outcome for each variant is also provided to aid in decision making.AvailabilityThe model and running code are available for free only for non-commercial users at https://www.github.com/aditya-88/intelli-ngs.

An Application of eHealth Technology Toward the Digitization of the Health Records of Older Patients With Cochlear Implants

Purpose Despite the current legislative indications toward the digitization of patient health records, 80% of health data are unstructured and in a format that cannot be used in electronic archives or in registries of diseases. An innovative automated system is here proposed to efficiently retrieve and digitize clinical information from original unstructured ear, nose, and throat (ENT) medical records, in order to reduce the manual workload in the retrieval and digitization process. Method The system, based on an eHealth technology named cognitive computing , interprets medical reports to transform unstructured clinical data (e.g., narrative text) into a structured digital format. The system has been tailored to handle the reports of aged cochlear implant (CI) patients by digitizing the information typically requested in electronic CI registries and by the current ENT/audiology guidelines. Results were obtained from the reports generated by an outpatient ENT care service from 52 older adult CI patients. Results The system allowed a quick and automated interpretation and retrieval of all the information required, such as the patient's medical history, risk factors, examination outcomes, communicative performances before and after CI implantation, and CI settings. The accuracy of the system in correctly interpreting and retrieving the above information from the original unstructured medical reports was very good (recall = 0.78; precision = 0.95). The system allowed to reduce the time needed to manually digitize the information from 20–30 min/report to only 20 s/report. Conclusion The proposed system is a viable solution for the automated digitization of unstructured health data as recommended by the ENT/audiology clinical best practices.

Synthetic Minority Oversampling and Smote Regularised Deep Autoencoders Neural Network Techniques for Fraud Prediction in Financial Payment Services

Frauds in Financial Payment Services are the most prevalent form of cybercrime. The increased growth in e-commerce and mobile payments in recent years is behind the rising incidence of fraud in financial payment services. According to "McKinsey, fraud losses throughout the world could be close to $44 billion by 2025." Every year, fraudulent card transactions causes billions of US Dollar of loss. To reduce these losses, designing effective fraud detection algorithms is essential, which depend on sophisticated machine learning methods to help investigators in fraud. For banks and financial institutions, therefore, fraud detection systems have gained excellent significance. Though the fake transactions are very low when compared to genuine transaction, care must be taken to predict it so that the financial institutions can maintain the customer integrity. As fraud is unlikely to occur compared to normal operations, we have the class imbalance problem. We applied Synthetic Minority Oversampling TEchnique (SMOTE) and the Ensemble of sampling methods(Balanced Random Forest Classifier, Balanced Bagging Classifier, Easy Ensemble Classifier, RUS Boost) to Ensemble machine learning algorithms Performance assessment using sensitivity, specificity, precision, ROC area. The purpose of this article is to analyze different predictive models to see how precise they are to detect whether a transaction is a standard payment or a fraud. Instead of misclassifying a real transaction as fraud, this model seeks to improve detection of fraud. We noted that the technique of Ensemble learning using Maximum voting detects the fraud better than other classifiers. Decision Tree Classifier, Logistic Regression, Balanced Bagging classifier is combined and the proposed algorithm is OptimizedEnsembleFD Algorithm. The sample size is increased and deep learning is applied .It is found that the proposed system Smote Regularised Deep Autoencoders (SRD Autoencoders) neural network performs better with good recall and accuracy for this large dataset.

Multi Sensor Image Matching using Super Symmetric Affinity Tensors based HyperGraph Matching

Image Matching technique is regularly on one of the main errands in numerous Photogrammetry and Remote Sensing applications. Based on multi-discipline, the approach of multiple sensor image matching is a novel one established which has vital application in military, civil, medicinal, and certain other domains. However, image matching approach faces numerous challenges, specifically in multi-sensor images where the images are gathered from the different sensor with different intensities, scales, and moments. Thus, a novel image matching approach is introduced in this paper using affinity tensor and HyperGraph Matching (HGM) technique that attempts to overcome certain drawbacks in matching and increases performance accuracy. Hypergraph matching techniques are employed using affinity tensors and consider supersymmetric property during construction. Graphs are constructed using graph theory for both sources, and target image and matching is done using third-order tensors. The experimental outcomes displayed that the proposed technique has good recall, precision, and positive accuracy values compared to the existing two descriptors based and tensor-based matching algorithms.

GeNet: Deep Representations for Metagenomics

We introduce GeNet, a method for shotgun metagenomic classification from raw DNA sequences that exploits the known hierarchical structure between labels for training. We provide a comparison with state-of-the-art methods Kraken and Centrifuge on datasets obtained from several sequencing technologies, in which dataset shift occurs. We show that GeNet obtains competitive precision and good recall, with orders of magnitude less memory requirements. Moreover, we show that a linear model trained on top of representations learned by GeNet achieves recall comparable to state-of-the-art methods on the aforementioned datasets, and achieves over 90% accuracy in a challenging pathogen detection problem. This provides evidence of the usefulness of the representations learned by GeNet for downstream biological tasks.

Freight car target detection in a complex background based on convolutional neural networks

Freight car target detection plays an important role in railway traffic safety, which typically depends on artificial observation or conventional machine learning, with insufficient accuracy and high demand for an observer's physical strength and image quality. Motivated by the recent advances of the convolutional neural network in object detection, this study investigates how deep neural networks can be applied in freight car target detection to better solve the aforementioned problems. We propose a novel two-training method for freight car target detection; the method includes general training and special training. In addition, online hard example mining and deformable convolutional network are introduced to select hard examples and extract better features for the special training stage to improve the problem of tiny target detection in poor images obtained from freight car target detection. The proposed methods are verified using experimental results based on three aspects, i.e. indexes, visualization, and speed. High accuracy can be achieved with good recall and acceptable speed for freight car target detection applications. Finally, we illustrate the utility of using such a model to test high robustness for changes in image quality and other target detection tasks with slight modification.