Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

General General

How the COVID-19 pandemic favored the adoption of digital technologies in healthcare: a systematic review of early scientific literature.

In Journal of medical Internet research ; h5-index 88.0

BACKGROUND : The COVID-19 pandemic is favoring the digital transition in many industries and in the society as a whole. Healthcare responded to the first phase of the pandemic through the rapid adoption of digital solutions and advanced technology tools.

OBJECTIVE : The aim of this study is to describe which digital solutions have been reported in the early scientific literature to mitigate the impact of COVID-19 on individuals and health systems.

METHODS : We conducted a systematic review of COVID-19 early literature (from January 1, 2020 to April 30, 2020) searching MEDLINE and medRxiv with terms considered adequate to find relevant literature on the use of digital technologies in response to the pandemic. We extracted study characteristics such as paper title, journal, publication date, and categorized the retrieved papers by type of technology, and patient needs addressed. We built a scoring rubric by cross-classifying the patient needs with the type of technology. We also extracted information and classified each technology reported by the selected articles according to healthcare system targets, grade of innovation, and scalability to other geographical areas.

RESULTS : The search identified 269 articles, of which 124 full-text articles were assessed and included in the review after screening. Of selected articles, most of them addressed the use of digital technologies for diagnosis, surveillance and prevention. We report that digital solutions and innovative technologies have mainly been proposed for the diagnosis of COVID-19. In particular, within the reviewed articles we identified numerous suggestions on the use of artificial-intelligence-powered tools for the diagnosis and screening of COVID-19. Digital technologies are useful also for prevention and surveillance measures, for example through contact-tracing apps or monitoring of internet searches and social media usage. Fewer scientific contributions address the use of digital technologies for lifestyle empowerment or patient engagement.

CONCLUSIONS : In the field of diagnosis, digital solutions that integrate with the traditional methods, such as AI-based diagnostic algorithms based both on imaging and/or clinical data, seem promising. As for surveillance, digital apps have already proven their effectiveness, but problems related to privacy and usability remain. For other patient needs, several solutions have been proposed using, for example, telemedicine or telehealth tools. These have long been available, but perhaps this historical moment could actually favor their definitive large-scale adoption. It is worth taking advantage of the push given by the crisis and important to keep track of the digital solutions proposed today to implement tomorrow's best practices and models of care, and to adopt at least some of the solutions proposed in the scientific literature, especially in those national health systems which in recent years proved to be particularly resistant to the digital transition.


Golinelli Davide, Boetto Erik, Carullo Gherardo, Nuzzolese Andrea Giovanni, Landini Maria Paola, Fantini Maria Pia


Pathology Pathology

PARENTing via Model-Agnostic Reinforcement Learning to Correct Pathological Behaviors in Data-to-Text Generation

ArXiv Preprint

In language generation models conditioned by structured data, the classical training via maximum likelihood almost always leads models to pick up on dataset divergence (i.e., hallucinations or omissions), and to incorporate them erroneously in their own generations at inference. In this work, we build ontop of previous Reinforcement Learning based approaches and show that a model-agnostic framework relying on the recently introduced PARENT metric is efficient at reducing both hallucinations and omissions. Evaluations on the widely used WikiBIO and WebNLG benchmarks demonstrate the effectiveness of this framework compared to state-of-the-art models.

Clément Rebuffel, Laure Soulier, Geoffrey Scoutheeten, Patrick Gallinari


Radiology Radiology

Eigenrank by committee: Von Neumann entropy based data subset selection and failure prediction for deep learning based medical image segmentation.

In Medical image analysis

Manual delineation of anatomy on existing images is the basis of developing deep learning algorithms for medical image segmentation. However, manual segmentation is tedious. It is also expensive because clinician effort is necessary to ensure correctness of delineation. Consequently most algorithm development is based on a tiny fraction of the vast amount of imaging data collected at a medical center. Thus, selection of a subset of images from hospital databases for manual delineation - so that algorithms trained on such data are accurate and tolerant to variation, becomes an important challenge. We address this challenge using a novel algorithm. The proposed algorithm named 'Eigenrank by Committee' (EBC) first computes the degree of disagreement between segmentations generated by each DL model in a committee. Then, it iteratively adds to the committee, a DL model trained on cases where the disagreement is maximal. The disagreement between segmentations is quantified by the maximum eigenvalue of a Dice coefficient disagreement matrix a measure closely related to the Von Neumann entropy. We use EBC for selecting data subsets for manual labeling from a larger database of spinal canal segmentations as well as intervertebral disk segmentations. U-Nets trained on these subsets are used to generate segmentations on the remaining data. Similar sized data subsets are also randomly sampled from the respective databases, and U-Nets are trained on these random subsets as well. We found that U-Nets trained using data subsets selected by EBC, generate segmentations with higher average Dice coefficients on the rest of the database than U-Nets trained using random sampling (p < 0.05 using t-tests comparing averages). Furthermore, U-Nets trained using data subsets selected by EBC generate segmentations with a distribution of Dice coefficients that demonstrate significantly (p < 0.05 using Bartlett's test) lower variance in comparison to U-Nets trained using random sampling for all datasets. We believe that this lower variance indicates that U-Nets trained with EBC are more robust than U-Nets trained with random sampling.

Gaonkar Bilwaj, Beckett Joel, Attiah Mark, Ahn Christine, Edwards Matthew, Wilson Bayard, Laiwalla Azim, Salehi Banafsheh, Yoo Bryan, Bui Alex A T, Macyszyn Luke


Active learing, Data subset selection, Deep learning, Failure deep learning

General General

Automatic myocardial infarction detection in contrast echocardiography based on polar residual network.

In Computer methods and programs in biomedicine

PURPOSE : Heart disease is one of the leading causes of death. Among patients with cardiovascular diseases, myocardial infarction (MI) is the main reason. Precise and timely identification of MI is significant for early treatment. Myocardial contrast echocardiography (MCE) is widely used for the detection of MI in clinic practice. However, existing clinical exam using MCE is subjective and highly operator dependent and time-consuming. Hence an automatic computer-aided MI detection in MCE is necessary to improve the diagnosis performance and decrease the workload of clinicians.

METHODS : In this study, a novel deep learning model, polar residual network (PResNet) is proposed to identify MI regions in MCE images which design a polar layer considering the ring shape of the myocardium. MCE images are fed into the PResNet and a newly defined polar layer is used to describe the myocardium with a ring shape. The whole polar images are evenly divided into several subsections and a residual network is improved to classify the subsection into normal and abnormal categories. Finally, the detection results are mapped back to the original image to illustrate the infarction regions' locations for the further process.

RESULTS : To evaluate the proposed PResNet, a dataset is constructed via performing MCE on five mice, which underwent the left anterior descending artery ligation and receive erythropoietin or saline injection, and the area variation fraction is manually annotated by an experienced expert as golden standards. The results demonstrate that the proposed PResNet model accomplishes high classification precisions with 99.6% and 98.7%, and 0.999 and 0.996 of AUC (area under the receiver operator curve) values on two different testing sets, respectively. Results suggest that the proposed model could enable accurate infarct detection and diagnosis of the MCE images.

CONCLUSION : Those efficiency gains highlight the powerful ability to describe and interpret the MCE images using the polar layer and residual network. The proposed PResNet might aid the clinicians in fast and accurate assessing the infarcted myocardium on MCE.

Guo Yanhui, Du Guo-Qing, Shen Wen-Qian, Du Chunlai, He Pei-Na, Siuly Siuly


Convolutional neural network, Deep learning, Infarct detection, Myocardial contrast echocardiography, Polar system

General General

Effector Biology of Biotrophic Plant Fungal Pathogens: Current Advances and Future Prospects.

In Microbiological research

The interaction of fungal pathogens with their host requires a novel invading mechanism and the presence of various virulence-associated components responsible for promoting the infection. The small secretory proteins, explicitly known as effector proteins, are one of the prime mechanisms of host manipulation utilized by the pathogen to disarm the host. Several effector proteins are known to translocate from fungus to the plant cell for host manipulation. Many fungal effectors have been identified using genomic, transcriptomic, and bioinformatics approaches. Most of the effector proteins are devoid of any conserved signatures, and their prediction based on sequence homology is very challenging, therefore by combining the sequence consensus based upon machine learning features, multiple tools have also been developed for predicting apoplastic and cytoplasmic effectors. Various post-genomics approaches like transcriptomics of virulent isolates have also been utilized for identifying active consortia of effectors. Significant progress has been made in understanding biotrophic effectors; however, most of it is underway due to their complex interaction with host and complicated recognition and signaling networks. This review discusses advances, and challenges in effector identification and highlighted various features of the potential effector proteins and approaches for understanding their genetics and strategies for regulation.

Jaswal Rajdeep, Kiran Kanti, Rajarammohan Sivasubramanian, Dubey Himanshu, Singh Pankaj Kumar, Sharma Yogesh, Deshmukh Rupesh, Sonah Humira, Gupta Naveen, Sharma T R


Plant fungal pathogens, biotrophs, effectors, genomics, host-pathogen interaction

General General

Associated Learning: Decomposing End-to-End Backpropagation Based on Autoencoders and Target Propagation.

In Neural computation

Backpropagation (BP) is the cornerstone of today's deep learning algorithms, but it is inefficient partially because of backward locking, which means updating the weights of one layer locks the weight updates in the other layers. Consequently, it is challenging to apply parallel computing or a pipeline structure to update the weights in different layers simultaneously. In this letter, we introduce a novel learning structure, associated learning (AL), that modularizes the network into smaller components, each of which has a local objective. Because the objectives are mutually independent, AL can learn the parameters in different layers independently and simultaneously, so it is feasible to apply a pipeline structure to improve the training throughput. Specifically, this pipeline structure improves the complexity of the training time from O ( n ) , which is the time complexity when using BP and stochastic gradient descent (SGD) for training, to O ( n + ) , where n is the number of training instances and is the number of hidden layers. Surprisingly, even though most of the parameters in AL do not directly interact with the target variable, training deep models by this method yields accuracies comparable to those from models trained using typical BP methods, in which all parameters are used to predict the target variable. Consequently, because of the scalability and the predictive power demonstrated in the experiments, AL deserves further study to determine the better hyperparameter settings, such as activation function selection, learning rate scheduling, and weight initialization, to accumulate experience, as we have done over the years with the typical BP method. In addition, perhaps our design can also inspire new network designs for deep learning. Our implementation is available at

Kao Yu-Wei, Chen Hung-Hsuan