Flash-based content addressable memory with L2 distance for memory-augmented neural network.

Memory-augmented neural network (MANN) has received increasing attention as a promising approach to achieve lifelong on-device learning, of which implementation of the explicit memory is vital. Content addressable memory (CAM) has been designed to accelerate the explicit memory by harnessing the in-memory-computing capability. In this work, a CAM cell with quadratic code is proposed, and a 1Mb Flash-based multi-bit CAM chip capable of computing Euclidean (L2) distance is fabricated. Compared with ternary CAM, the latency and energy are significantly reduced by 5.3- and 46.6-fold, respectively, for the MANN on Omniglot dataset. Besides, the recognition accuracy has slight degradation (<1%) even after baking for 105 s at 200°C, demonstrating the robustness to environmental disturbance. Performance evaluation indicates a reduction of 471-fold in latency and 1267-fold in energy compared with GPU for search operation. The proposed robust and energy-efficient CAM provides a promising solution to implement lifelong on-device machine intelligence.

A 3D-printed neuromorphic humanoid hand for grasping unknown objects.

Compared with conventional von Neumann's architecture-based processors, neuromorphic systems provide energy-saving in-memory computing. We present here a 3D neuromorphic humanoid hand designed for providing an artificial unconscious response based on training. The neuromorphic humanoid hand system mimics the reflex arc for a quick response by managing complex spatiotemporal information. A 3D structural humanoid hand is first integrated with 3D-printed pressure sensors and a portable neuromorphic device that was fabricated by the multi-axis robot 3D printing technology. The 3D neuromorphic robot hand provides bioinspired signal perception, including detection, signal transmission, and signal processing, together with the biomimetic reflex arc function, allowing it to hold an unknown object with an automatically increased gripping force without a conventional controlling processor. The proposed system offers a new approach for realizing an unconscious response with an artificially intelligent robot.

Neural network-based cooling design for high-performance processors.

Ultra-high chip power densities that are expected to surpass 1-2kW/cm2 in future high-performance systems cannot be easily handled by conventional cooling methods. Various emerging cooling methods, such as liquid cooling via microchannels, thermoelectric coolers (TECs), two-phase vapor chambers, and hybrid cooling options have been designed to efficiently remove heat from high-performance processors. However, selecting the optimal cooling solution for a given chip and determining the optimal cooling parameters for that solution to achieve high efficiency are open problems. These problems are, in fact, computationally expensive because of the massive space of possible solutions. To address this design challenge, this article introduces a deep learning-based cooling design optimization flow that rapidly and accurately converges to the optimal cooling solution as well as the optimal cooling parameters for a given chip floorplan and its power profile.

Use of Digital Technology among Adolescents Attending Schools in Bissau, Guinea-Bissau.

Digital technology plays an important role in achieving many of the Sustainable Development Goals. However, access is uneven, with 80% of those in high-income countries being online compared to 20% of those in the 47 least developed countries. This study aimed to describe and analyse adolescents' access to and usage of digital technology in Guinea-Bissau and its implications. In June 2017, a survey with a locally adapted Planet Youth questionnaire was implemented in the capital, Bissau, whereby classes in 16 secondary schools were surveyed on a variety of issues. In total, 2039 randomly selected students participated; the survey included ten questions specifically on the access to and use of digital technology. Half of the respondents had access to desktop/laptops, and one-third used mobile internet daily; about two-thirds had an experience of social media. Explanatory variables included educational institution, parental education, economic situation, and gender. Furthermore, students' experience of social media was significantly linked to bullying, anxiety, depression, smoking and alcohol consumption. Many adolescents in Bissau have no experience of using digital technology, including for schoolwork. Access improvements are necessary so that young Bissau-Guineans are not to be left behind in developing their capabilities and can benefit from proficiency in the use of digital technologies. At the same time, potential harmful usage of the media requires the implementation of preventive measures.

Plasticity and Adaptation in Neuromorphic Biohybrid Systems.

Neuromorphic systems take inspiration from the principles of biological information processing to form hardware platforms that enable the large-scale implementation of neural networks. The recent years have seen both advances in the theoretical aspects of spiking neural networks for their use in classification and control tasks and a progress in electrophysiological methods that is pushing the frontiers of intelligent neural interfacing and signal processing technologies. At the forefront of these new technologies, artificial and biological neural networks are tightly coupled, offering a novel "biohybrid" experimental framework for engineers and neurophysiologists. Indeed, biohybrid systems can constitute a new class of neuroprostheses opening important perspectives in the treatment of neurological disorders. Moreover, the use of biologically plausible learning rules allows forming an overall fault-tolerant system of co-developing subsystems. To identify opportunities and challenges in neuromorphic biohybrid systems, we discuss the field from the perspectives of neurobiology, computational neuroscience, and neuromorphic engineering.

Sci-Thur AM: Planning - 11: The impact of distributed calculation framework settings on plan calculation time.

Some treatment planning system can divide a treatment plan calculation into multiple threads and allow both local and network computing resources to perform the calculation concurrently, which significantly reduces the calculation time for a calculation-demanding planning such as Volumetric Modulated Arc Therapy (VMAT) or electron Monte Carlo (eMC). This study tested in Eclipse (Varian, V10.0.39) the impact of Distributed Calculation Framework (DCF, V10.0.0.757) settings on calculation time in a planning environment that consists of 20 workstations with 8 core processors and 16GB RAMs installed on most of them. It is found that for an arc plan increasing the control point field parallelization factor reduces the total calculation time at beginning but lengthens the total calculation time after a certain level as a result of data sending time increase. Further increasing the factor may cause a serious net work traffic or even failure of a calculation. For an eMC plan the calculation time decreases monotonously with the increase of Monte carlo field parallelization factor, and the data sending time is insignificant compared to the calculation time. Increasing the local servant numbers reduces the data sending time but raises the calculation time for arc and eMC plans. The calculation time increment is more and more significant with the increase of local servants. The optimal DCF setting for a facility depends on the total number of calculation workstations available, the hardware configuration of the workstations, and the data transfer rate of the network. No conflict of interest exists in the study.

SU-E-T-213: Development of a Web Wrapper to Facilitate Radiotherapy Research.

PURPOSE: Researchers write many computer programs with unique implementations, usually requiring a great amount of effort for other researchers to learn how to install, configure, and use. Some programs require specialized hardware platforms such as GPU workstation or CPU cluster, which may not readily available for many researchers. This work develops a general web platform to 'wrap' radiotherapy software tools into a user friendly, browser-based interface. METHODS: We developed a web wrapper based on existing technologies (e.g. HTML5, JavaScript, PHP, Python, XML) to interface with command line-based research tools. This wrapper enables users to easily perform various tasks in any modern web browser, while underlying tools are launched remotely. Visitors can upload data, configure settings, process data remotely, then view, share, and download results with minimal effort. This web wrapper is developer friendly; new tools are easily integrated by editing XML configuration files. RESULTS: As a test case, we have successfully wrapped a set of command line tools, developed by our group, into a single web app, providing fluence map generation, CT image processing, and GPU-based Monte Carlo (MC) dose calculation. The result is a web-based quality assurance tool. With this tool, users can upload compressed DICOM-RT files, recompute dose using the MC method, and evaluate the results by viewing dose distribution, 3D gamma index distribution and DVH curves. The entire work-flow can be completed within 2 minutes provided users have a reasonable Internet connection speed. CONCLUSIONS: We have developed an web wrapper to increase the accessibility of radiotherapy tools and reduce users' learning curve through a friendly web-based interface. This work also allows quick and easy deployment and distribution of software tools developed by researchers to the whole community.

MO-D-213CD-01: Cartesian Methods for Rapid Time-Resolved MR Angiography.

The physics of the MR image formation fundamentally trades off spatial resolution with temporal resolution. Time spent in acquiring data for the second image of a time series can alternatively be spent in sampling higher spatial frequencies for the first image to improve its spatial resolution. Historically this tradeoff has been addressed by making the k-space sampling rate high, such as with very short repetition times, and with methods such as view sharing in which only a portion of k-space is updated from one image to the next in a time series. Over a decade ago the method of parallel acquisition was proposed in which the signals detected by the individual elements comprising a multi-element receiver coil are used to provide further spatial discrimination and reduce acquisition time. These approaches include those based in image space (SENSE) or in k- space (SMASH, GRAPPA). In the last decade these methods have been integrated in contrast-enhanced MRA (CE-MRA) to provide a radical improvement in performance. CE-MRA is an application particularly well suited to these methods. The general desire for MRA images to be three-dimensional allows the use of 2D implementation of parallel acquisition, generally much more robust than 1D implementation. Also, the SNR loss associated with parallel acquisition is tempered in CE-MRA because high, arterial-phase signal is sampled throughout the data acquisition. Cartesian MR data acquisition, performed along a rectilinear sampling pattern in k-space, offers specific advantages in relative ease of implementation of 2D parallel acquisition and in "freezing" the status of the time-varying object at a specific timepoint by use of centric view ordering. This presentation will provide a review of these methods and how they have been effectively developed and integrated within the last decade for improved time-resolved MRA. Cartesian k-space sampling patterns can now be quickly selected on a patient- and anatomy-specific basis for optimum acceleration. Receiver coil arrays have been adapted to allow up to 20× reduction in the number of k-space points sampled for a given spatial resolution. Reconstruction hardware now allows generation of 3D images within only hundreds of msec after data acquisition, permitting real-time generation of diagnostic quality images and their use in interactively guiding other processes. LEARNING OBJECTIVES: 1. Understand recently developed physics techniques which have allowed a 20x improvement in the speed of data acquisition for MR angiography 2. Understand how Cartesian sampling of k-space facilitates the practical and effective implementation of these techniques 3. Show how contemporary implementation of these physics techniques has provided a significant improvement in MRA image quality over the last decade.

SU-E-J-172: Development of a Video Guided Real-Time Patient Motion Monitoring System for Helical Tomotherpay.

PURPOSE: We developed a video image-guided real-time patient motion monitoring system for helical Tomotherapy (VGRPM-Tomo), and its clinical utility was evaluated using a motion phantom. METHODS: The VGRPM-Tomo consisted of three components: an image acquisition device consisting of two PC-cams, a main control computer with a radiation signal controller and warning system, and patient motion analysis software, which was developed in house. The system was designed for synchronization with a beam on/off trigger signal to limit operation during treatment time only and to enable system automation. In order to detect the patient motion while the couch is moving into the gantry, a reference image, which continuously updated its background by exponential weighting filter (EWF), is compared with subsequent live images using the real-time frame difference-based analysis software. When the error range exceeds the set criteria (δ_movement) due to patient movement, a warning message is generated in the form of light and sound. The described procedure repeats automatically for each patient. A motion phantom, which operates by moving a distance of 0.1, 0.2, 0.5, and 1.0 cm for 1 and 2 sec, respectively, was used to evaluate the system performance at maximum couch speed (0.196 cm/sec) in a Helical Tomotherapy (HD, Hi-art, Tomotherapy, USA). We measured the optimal EWF factor (a) and δ_movement, which is the minimum distance that can be detected with this system, and the response time of the whole system. RESULTS: The optimal a for clinical use ranged from 0.85 to 0.9. The system was able to detect phantom motion as small as 0.2 cm with tight δ_movement, 0.1% total number of pixels in the reference image. The measured response time of the whole system was 0.1 sec. CONCLUSIONS: The VGRPM-tomo can contribute to reduction of treatment error caused by the motion of patients and increase the accuracy of treatment dose delivery in HD. This work was supported by the Technology Innovation Program, 10040362, Development of an integrated management solution for radiation therapy funded by the Ministry of Knowledge Economy (MKE, Korea). This idea is protected by a Korean patent (patent no. 10-1007367).

WE-E-217A-02: Methodologies for Evaluation of Standalone CAD System Performance.

Standalone performance evaluation of a CAD system provides information about the abnormality detection or classification performance of the computerized system alone. Although the performance of the reader with CAD is the final step in CAD system assessment, standalone performance evaluation is an important component for several reasons: First, standalone evaluation informs the reader about the performance level of the CAD system and may have an impact on how the reader uses the system. Second, it provides essential information to the system designer for algorithm optimization during system development. Third, standalone evaluation can provide a detailed description of algorithm performance (e.g., on subgroups of the population) because a larger data set with more samples from different subgroups can be included in standalone studies compared to reader studies. Proper standalone evaluation of a CAD system involves a number of key components, some of which are shared with the assessment of reader performance with CAD. These include (1) selection of a test data set that allows performance assessment with little or no bias and acceptable uncertainty; (2) a reference standard that indicates disease status as well as the location and extent of disease; (3) a clearly defined method for labeling each CAD mark as a true-positive or false-positive; and (4) a properly selected set of metrics to summarize the accuracy of the computer marks and their corresponding scores. In this lecture, we will discuss various approaches for the key components of standalone CAD performance evaluation listed above, and present some of the recommendations and opinions from the AAPM CAD subcommittee on these issues. Learning Objectives 1. Identify basic components and metrics in the assessment of standalone CAD systems 2. Understand how each component may affect the assessed performance 3. Learn about AAPM CAD subcommittee's opinions and recommendations on factors and metrics related to the evaluation of standalone CAD system performance.