Object Detection and Instance Segmentation in Chest X-rays for Tuberculosis Screening

Tuberculosis (TB) is a highly contagious disease leading to the deaths of approximately 2 million people annually. TB primarily affects the lungs and is spread through the air when people cough, sneeze, or spit. Providing healthcare professionals with better information, at a faster pace, is essential for combating this disease, especially in Low and Middle Income Countries (LMICs) with resource-constrained health systems. In this paper we describe how using convolution neural networks (CNNs) with an object level annotated dataset of chest X-rays (CXRs) allows us to identify the location of pulmonary issues indicative of TB. We compare the performance of Faster R-nobreakdash-CNN, Mask R-nobreakdash-CNN, Cascade versions of each, and SOLOv2, demonstrating reasonable results with a small dataset. We present a method to reduce the false positive rate by comparing the location of a detected object with the known location of areas where the detected class is likely to occur in the lung. Our results show that object detection and instance segmentation of CXRs can be achieved with a dataset of high-quality, object level annotations, and could be used as part of an automated TB screening process. This work has the potential to improve the speed of TB diagnosis in LMICs, if properly integrated into the healthcare system and adapted to existing clinical workflows and local regulations.

Estimating Number and Dwell Time of Visitors in a Large-scale Indoor Space from HVAC Data and Its Evaluation in Different Seasons

The aim of this paper is to estimate the number and dwell time of visitors in a large-scale indoor space or room with a common heating-ventilation-air conditioning (HVAC) system that includes sensors for CO2 and indoor temperature in any season. Previous studies tried to estimate the number and dwell time of visitors from CO2 concentration in small rooms with or without a HVAC system. However, in a large-scale indoor space with large air-conditioning and ventilation systems, the number and dwell time of visitors are difficult to estimate for three reasons: 1) CO2 concentration changes much more slowly than the number and dwell time of visitors and with a delay, 2) the difference in changes is affected by the amount of ventilation, and 3) this difference may be affected by operation of HVAC, which is affected by seasonal climate. To solve these problems, we proposed partial modeling with a variable time window. This method can make a partial estimation model that automatically corresponds to differences in the change speed between two variables: visitors and CO2 concentration. We demonstrate the effectiveness of our proposal using measured sensor data in summer, fall, and winter to clarify its feasibility in different seasons in Japan.

Behavioral Link Analytics on Heterogeneous Human Interaction Networks

For designing and modeling Artificial Intelligence (AI) systems in the area of human-machine interaction, suitable approaches for user modeling are important in order to both capture user characteristics. Using multimodal data, this can be performed from various perspectives. Specifically, for modeling user interactions in human interaction networks, appropriate approaches for capturing those interactions, as well as to analyze them in order to extract meaningful patterns are important. Specifically, for modeling user behavior for the respective AI systems, we can make use of diverse heterogeneous data sources. This paper investigates face-to-face as well as socio-spatial interaction networks for modeling user interactions from three perspectives: We analyze preferences and perceptions of human social interactions in relation to the interactions observed using wearable sensors, i. e., face-to-face as well as socio-spatial interactions fo the respective actors. For that, we investigate the correspondence of according networks, in order to identify conformance, exceptions, and anomalies. The analysis is performed on a real-world dataset capturing networks of proximity interactions coupled with self-report questionnaires about preferences and perception of those interactions. The different networks, and according perspectives then provide different options for user modeling and integration into AI systems modeling such user behavior.

Measuring Group Personality with Swarm AI

The aggregation of individual personality assessments to predict team performance is widely accepted in management theory but has significant limitations: the isolated nature of individual personality surveys fails to capture much of the team dynamics that drive realworld team performance. Artificial Swarm Intelligence (ASI)—a technology that enables networked teams to think together in real-time and answer questions as a unified system—promises a solution to these limitations by enabling teams to collectively complete a personality assessment, whereby the team uses ASI to converge upon answers that best represent the group’s disposition. In the present study, the group personality of 94 small teams was assessed by having teams take a standard Big Five Inventory (BFI) assessment both as individuals, and as a realtime system enabled by an ASI technology known as Swarm AI. The predictive accuracy of each personality assessment method was assessed by correlating the BFI personality traits to a range of real-world performance metrics. The results showed that assessments of personality generated using Swarm AI were far more predictive of team performance than the traditional aggregation methods, showing at least a 91.8% increase in average correlation with the measured outcome variables, and in no case showing a significant decrease in predictive performance. This suggests that Swarm AI technology may be used as a highly effective team personality assessment tool that more accurately predicts future team performance than traditional survey approaches.

Generating Metrically Accurate Homeric Poetry with Recurrent Neural Networks

We investigate the generation of metrically accurate Homeric poetry using recurrent neural networks (RNN). We assess two models: a basic encoder-decoder RNN and the hierarchical recurrent encoderdecoder model (HRED). We assess the quality of the generated lines of poetry using quantitative metrical analysis and expert evaluation. This evaluation reveals that while the basic encoder-decoder is able to capture complex poetic meter, it under performs in terms of semantic coherence. The HRED model, however, produces more semantically coherent lines of poetry but is unable to capture the meter. Our research highlights the importance of expert evaluation and suggests that future research should focus on encoder-decoder models that balance various types of input – both immediate and long-range.