Search

Recurrent neural networks (RNNs) are a class of neural networks that are naturally suited to processing time-series data and other sequential data. Here we introduce recurrent neural networks as an extension to feedforward networks, in order to allow the processing of variable-length (or even infinite-length) sequences, and some of the most popular recurrent architectures in use, including long short-term memory (LSTM) and gated recurrent units (GRUs). In addition, various aspects surrounding RNNs are discussed in detail, including various probabilistic models that are often realized using RNNs and various applications of RNNs that have appeared within the MICCAI community.

Automatically segmenting and classifying surgical activities is an important prerequisite to providing automated, targeted assessment and feedback during surgical training. Prior work has focused almost exclusively on recognizing gestures, or short, atomic units of activity such as pushing needle through tissue, whereas we also focus on recognizing higher-level maneuvers, such as suture throw. Maneuvers exhibit more complexity and variability than the gestures from which they are composed, however working at this granularity has the benefit of being consistent with existing training curricula.Prior work has focused on hidden Markov model and conditional-random-field-based methods, which typically leverage unary terms that are local in time and linear in model parameters. Because maneuvers are governed by long-term, nonlinear dynamics, we argue that the more expressive unary terms offered by recurrent neural networks (RNNs) are better suited for this task. Four RNN architectures are compared for recognizing activities from kinematics: simple RNNs, long short-term memory, gated recurrent units, and mixed history RNNs. We report performance in terms of error rate and edit distance, and we use a functional analysis-of-variance framework to assess hyperparameter sensitivity for each architecture.We obtain state-of-the-art performance for both maneuver recognition from kinematics (4 maneuvers; error rate of [Formula: see text]; normalized edit distance of [Formula: see text]) and gesture recognition from kinematics (10 gestures; error rate of [Formula: see text]; normalized edit distance of [Formula: see text]).Automated maneuver recognition is feasible with RNNs, an exciting result which offers the opportunity to provide targeted assessment and feedback at a higher level of granularity. In addition, we show that multiple hyperparameters are important for achieving good performance, and our hyperparameter analysis serves to aid future work in RNN-based activity recognition.

One-shot pose estimation for tasks such as body joint localization, camera pose estimation, and object tracking are generally noisy, and temporal filters have been extensively used for regularization. One of the most widely-used methods is the Kalman filter, which is both extremely simple and general. However, Kalman filters require a motion model and measurement model to be specified a priori, which burdens the modeler and simultaneously demands that we use explicit models that are often only crude approximations of reality. For example, in the pose-estimation tasks mentioned above, it is common to use motion models that assume constant velocity or constant acceleration, and we believe that these simplified representations are severely inhibitive. In this work, we propose to instead learn rich, dynamic representations of the motion and noise models. In particular, we propose learning these models from data using long shortterm memory, which allows representations that depend on all previous observations and all previous states. We evaluate our method using three of the most popular pose estimation tasks in computer vision, and in all cases we obtain state-of-the-art performance.

Many military and civilian applications depend on the ability to remotely sense chemical agent (CA) clouds, from detecting small but lethal concentrations of chemical warfare agents (CWAs) to mapping plumes in the aftermath of natural disasters. Hyperspectral imaging sensors, in the long-wave infrared (IR) (LWIR) part of the spectrum, are particularly well suited for these gas-sensing tasks. Their wide fields of view allow many square kilometers to be imaged almost simultaneously, and their high spectral resolutions often allow the detection of chemical clouds even when they are optically thin (or nearly transparent).

We apply recurrent neural networks to the task of recognizing surgical activities from robot kinematics. Prior work in this area focuses on recognizing short, low-level activities, or gestures, and has been based on variants of hidden Markov models and conditional random fields. In contrast, we work on recognizing both gestures and longer, higher-level activites, or maneuvers, and we model the mapping from kinematics to gestures/maneuvers with recurrent neural networks. To our knowledge, we are the first to apply recurrent neural networks to this task. Using a single model and a single set of hyperparameters, we match state-of-the-art performance for gesture recognition and advance state-of-the-art performance for maneuver recognition, in terms of both accuracy and edit distance. Code is available at https://github.com/rdipietro/miccai-2016-surgical-activity-rec.

The utility and promise of magnetic nanoparticles (MagNPs) for biomedicine rely heavily on accurate determination of the particle diameter attributes. While the average functional size and size distribution of the magnetic nanoparticles directly impact the implementation and optimization of nanobiotechnology applications in which they are employed, the determination of these attributes using electron microscopy techniques can be time-consuming and misrepresentative of the full nanoparticle population. In this work the average particle diameter and distribution of an ensemble of Fe 3 O 4 ferrimagnetic nanoparticles are determined solely from temperature-dependent magnetization measurements; the results compare favorably to those obtained from extensive electron microscopy observations. The attributes of a population of biocompatible Fe 3 O 4 nanoparticles synthesized by a thermal decomposition method are obtained from quantitative evaluation of a model that incorporates the distribution of superparamagnetic blocking temperatures represented through thermomagnetization data. The average size and size distributions are determined from magnetization data via temperature-dependent zero-field-cooled magnetization. The current work is unique from existing approaches based on magnetic measurement for the characterization of a nanoparticle ensemble as it provides both the average particle size as well as the particle size distribution.

Many military and civilian applications depend on the ability to remotely sense chemical clouds using hyperspectral imagers, from detecting small but lethal concentrations of chemical warfare agents to mapping plumes in the aftermath of natural disasters. Real-time operation is critical in these applications but becomes diffcult to achieve as the number of chemicals we search for increases. In this paper, we present efficient CPU and GPU implementations of matched-filter based algorithms so that real-time operation can be maintained with higher chemical-signature counts. The optimized C++ implementations show between 3x and 9x speedup over vectorized MATLAB implementations.

Chemical cloud detection using long-wave infrared (LWIR) hyperspectral-imaging sensors has many civilian and military applications, including chemical warfare threat mitigation, environmental monitoring, and emergency response. Current capabilities are limited by variation in background clutter as opposed to the physics of photon detection, and this makes the statistical characterization of clutter and clutter-induced false alarms essential to the design of practical systems. In this exploratory work, we use hyperspectral data collected both on the ground and in the air to spectrally and spatially characterize false alarms. Focusing on two widely-used detectors, the matched filter (MF) and the adaptive cosine estimator (ACE), we compare empirical false-alarm rates to their theoretical counterparts - detector output under Gaussian, t and t-mixture distributed data - and show that these models often underestimate false-alarm rates. Next, we threshold real detection maps and show that true detections and false alarms often exhibit very different spatial behavior. To exploit this difference and understand how spatial processing affects performance, the spatial behavior of false alarms must be understood. We take a first step in this direction by showing that, although the behavior may `look' quite random, it is not well captured by the complete-spatial-randomness model. Finally, we describe how our findings impact the design of real detection systems.

One of the fundamental challenges for a hyperspectral imaging surveillance system is the detection of sub-pixel objects in background clutter. The background surrounding the object, which acts as interference, provides the major obstacle to successful detection. Two additional limiting factors are the spectral variabilities of the background and the object to be detected. In this paper, we evaluate the performance of detection algorithms for sub-pixel objects using a replacement signal model, where the spectral variability is modeled by multivariate normal distributions. The detection algorithms considered are the classical matched filter, the matched filter with false alarm mitigation, the mixture tuned matched filter and the finite target matched filter. These algorithms are compared using simulated and actual hyperspectral imaging data.

One of the fundamental challenges for a hyperspectral imaging surveillance system is the detection of sub-pixel objects in background clutter. The background surrounding the object, which acts as interference, provides the major obstacle to successful detection. One algorithm that is widely used in hyperspectral detection and successfully suppresses the background in many situations is the matched filter detector. However, the matched filter also produces false alarms in many situations. We use three simple and well-established concepts-the target-background replacement model, the matched filter, and Mahalanobis distance-to develop the matched filter with false alarm mitigation (MF-FAM), a dual-threshold detector capable of eliminating many matched filter false alarms. We compare this algorithm to the mixture tuned matched filter (MTMF), a popular approach to matched filter false alarm mitigation found in the ENVI® software environment. The two algorithms are shown to produce nearly identical results using real hyperspectral data, but the MF-FAM is shown to be operationally, computationally, and theoretically simpler than the MTMF.

Large low field magnetoresistance (LFMR) of about 28% is observed in La0.67Sr0.33MnO3 nanowires with 80 nm in diameter at T=300 K. A gradual decrease in the LFMR has been found with increase in wire diameter. The LFMR drops to zero for wires above 280 nm in diameter. The nanowires are grown by means of electrospinning process and exhibit distorted orthorhombic crystal structure. The large LFMR is considered as a grain boundary effect as observed in several perovskite systems. The large LFMR observed in these manganites with reduced dimensions may be useful for room temperature device applications.

Thermomagnetic measurements are used to obtain the size distribution and anisotropy of magnetic nanoparticles. An analytical transformation method is described which utilizes temperature-dependent zero-field cooling (ZFC) magnetization data to provide a quantitative measurement of the average diameter and relative abundance of superparamagnetic nanoparticles. Applying this method to self-assembled MnAs nanoparticles in MnAs-GaAs composite films reveals a log-normal size distribution and reduced anisotropy for nanoparticles compared to bulk materials. This analytical technique holds promise for rapid assessment of the size distribution of an ensemble of superparamagnetic nanoparticles., Comment: Correction Appl. Phys. Lett. 98, 216103 (2011)

Boron-incorporated Mn5SiC nanowires were grown using chemical vapor deposition method. The nanowire cluster exhibits magnetic hysteresis loops at room temperature and the strength of the magnetic behavior depends on the concentration of the boron incorporation. Mn5SiC nanowire-based devices exhibit spin dependent transport properties which shows significant changes with boron content. Large magnetoresistance is observed in lightly boron-incorporated nanowire devices and it decreases with increase in boron content.