Using conditional generative adversarial networks to reduce the effects of latency in robotic telesurgery.

The introduction of surgical robots brought about advancements in surgical procedures. The applications of remote telesurgery range from building medical clinics in underprivileged areas, to placing robots abroad in military hot-spots where accessibility and diversity of medical experience may be limited. Poor wireless connectivity may result in a prolonged delay, referred to as latency, between a surgeon's input and action which a robot takes. In surgery, any micro-delay can injure a patient severely and, in some cases, result in fatality. One way to increase safety is to mitigate the effects of latency using deep learning aided computer vision. While the current surgical robots use calibrated sensors to measure the position of the arms and tools, in this work, we present a purely optical approach that provides a measurement of the tool position in relation to the patient's tissues. This research aimed to produce a neural network that allowed a robot to detect its own mechanical manipulator arms. A conditional generative adversarial network (cGAN) was trained on 1107 frames of a mock gastrointestinal robotic surgery from the 2015 EndoVis Instrument Challenge and corresponding hand-drawn labels for each frame. When run on new testing data, the network generated near-perfect labels of the input images which were visually consistent with the hand-drawn labels and was able to do this in 299 ms. These accurately generated labels can then be used as simplified identifiers for the robot to track its own controlled tools. These results show potential for conditional GANs as a reaction mechanism, such that the robot can detect when its arms move outside the operating area in a patient. This system allows for more accurate monitoring of the position of surgical instruments in relation to the patient's tissue, increasing safety measures that are integral to successful telesurgery systems.

Gender Perception From Gait: A Comparison Between Biological, Biomimetic and Non-biomimetic Learning Paradigms.

This paper explores in parallel the underlying mechanisms in human perception of biological motion and the best approaches for automatic classification of gait. The experiments tested three different learning paradigms, namely, biological, biomimetic, and non-biomimetic models for gender identification from human gait. Psychophysical experiments with twenty-one observers were conducted along with computational experiments without applying any gender specific modifications to the models or the stimuli. Results demonstrate the utilization of a generic memory based learning system in humans for gait perception, thus reducing ambiguity between two opposing learning systems proposed for biological motion perception. Results also support the biomimetic nature of memory based artificial neural networks (ANN) in their ability to emulate biological neural networks, as opposed to non-biomimetic models. In addition, the comparison between biological and computational learning approaches establishes a memory based biomimetic model as the best candidate for a generic artificial gait classifier (83% accuracy, p < 0.001), compared to human observers (66%, p < 0.005) or non-biomimetic models (83%, p < 0.001) while adhering to human-like sensitivity to gender identification, promising potential for application of the model in any given non-gender based gait perception objective with superhuman performance.

Predicting Binding from Screening Assays with Transformer Network Embeddings.

Cheminformatics aims to assist in chemistry applications that depend on molecular interactions, structural characteristics, and functional properties. The arrival of deep learning and the abundance of easily accessible chemical data from repositories like PubChem have enabled advancements in computer-aided drug discovery. Virtual high-throughput screening (vHTS) is one such technique that integrates chemical domain knowledge to perform in silico biomolecular simulations, but prediction of binding affinity is restricted due to limited availability of ground-truth binding assay results. Here, text representations of 83 000 000 molecules are leveraged to perform single-target binding affinity prediction directly on the outcome of screening assays. The embedding of an end-to-end transformer neural network, trained to encode the structural characteristics of a molecule via a text-based translation task, is repurposed through transfer learning to classify binding affinity to single targets with few known binding compounds. We quantify the observed increase in AUC on binding prediction tasks between classifiers trained on the translation embedding versus those using an untrained embedding. Visualization of the embedding space reveals organization of structural and functional properties that aid binding prediction. The pretrained transformer, data, and associated software to extract embeddings are made publicly available at https://github.com/mpcrlab/MolecularTransformerEmbeddings.

Machine-learning approaches to substance-abuse research: emerging trends and their implications.

PURPOSE OF REVIEW: To provide an accessible overview of some of the most recent trends in the application of machine learning to the field of substance use disorders and their implications for future research and practice. RECENT FINDINGS: Machine-learning (ML) techniques have recently been applied to substance use disorder (SUD) data for multiple predictive applications including detecting current abuse, assessing future risk and predicting treatment success. These models cover a wide range of machine-learning techniques and data types including physiological measures, longitudinal surveys, treatment outcomes, national surveys, medical records and social media. SUMMARY: The application of machine-learning models to substance use disorder data shows significant promise, with some use cases and data types showing high predictive accuracy, particularly for models of physiological and behavioral measures for predicting current substance use, portending potential clinical diagnostic applications; however, these results are uneven, with some models performing poorly or at chance, a limitation likely reflecting insufficient data and/or weak validation methods. The field will likely benefit from larger and more multimodal datasets, greater standardization of data recording and rigorous testing protocols as well as greater use of modern deep neural network models applied to multimodal unstructured datasets.