Automatic Root Length Estimation from Images Acquired In Situ without Segmentation.

Image-based root phenotyping technologies, including the minirhizotron (MR), have expanded our understanding of the in situ root responses to changing environmental conditions. The conventional manual methods used to analyze MR images are time-consuming, limiting their implementation. This study presents an adaptation of our previously developed convolutional neural network-based models to estimate the total (cumulative) root length (TRL) per MR image without requiring segmentation. Training data were derived from manual annotations in Rootfly, commonly used software for MR image analysis. We compared TRL estimation with 2 models, a regression-based model and a detection-based model that detects the annotated points along the roots. Notably, the detection-based model can assist in examining human annotations by providing a visual inspection of roots in MR images. The models were trained and tested with 4,015 images acquired using 2 MR system types (manual and automated) and from 4 crop species (corn, pepper, melon, and tomato) grown under various abiotic stresses. These datasets are made publicly available as part of this publication. The coefficients of determination (R2), between the measurements made using Rootfly and the suggested TRL estimation models were 0.929 to 0.986 for the main datasets, demonstrating that this tool is accurate and robust. Additional analyses were conducted to examine the effects of (a) the data acquisition system and thus the image quality on the models' performance, (b) automated differentiation between images with and without roots, and (c) the use of the transfer learning technique. These approaches can support precision agriculture by providing real-time root growth information.

Medical concept embedding of real-valued electronic health records with application to inflammatory bowel disease.

Deep learning approaches are gradually being applied to electronic health record (EHR) data, but they fail to incorporate medical diagnosis codes and real-valued laboratory tests into a single input sequence for temporal modeling. Therefore, the modeling misses the existing medical interrelations among codes and lab test results that should be exploited to promote early disease detection. To find connections between past diagnoses, represented by medical codes, and real-valued laboratory tests, in order to exploit the full potential of the EHR in medical diagnosis, we present a novel method to embed the two sources of data into a recurrent neural network. Experimenting with a database of Crohn's disease (CD), a type of inflammatory bowel disease, patients and their controls (~1:2.2), we show that the introduction of lab test results improves the network's predictive performance more than the introduction of past diagnoses but also, surprisingly, more than when both are combined. In addition, using bootstrapping, we generalize the analysis of the imbalanced database to a medical condition that simulates real-life prevalence of a high-risk CD group of first-degree relatives with results that make our embedding method ready to screen this group in the population.

Deep Convolutional Tables: Deep Learning Without Convolutions.

We propose a novel formulation of deep networks that do not use dot-product neurons and rely on a hierarchy of voting tables instead, denoted as convolutional tables (CTs), to enable accelerated CPU-based inference. Convolutional layers are the most time-consuming bottleneck in contemporary deep learning techniques, severely limiting their use in the Internet of Things and CPU-based devices. The proposed CT performs a fern operation at each image location: it encodes the location environment into a binary index and uses the index to retrieve the desired local output from a table. The results of multiple tables are combined to derive the final output. The computational complexity of a CT transformation is independent of the patch (filter) size and grows gracefully with the number of channels, outperforming comparable convolutional layers. It is shown to have a better capacity:compute ratio than dot-product neurons, and that deep CT networks exhibit a universal approximation property similar to neural networks. As the transformation involves computing discrete indices, we derive a soft relaxation and gradient-based approach for training the CT hierarchy. Deep CT networks have been experimentally shown to have accuracy comparable to that of CNNs of similar architectures. In the low-compute regime, they enable an error:speed tradeoff superior to alternative efficient CNN architectures.

SIGN: Statistical Inference Graphs Based on Probabilistic Network Activity Interpretation.

Convolutional neural networks (CNNs) have achieved superior accuracy in many visual-related tasks. However, the inference process through a CNN's intermediate layers is opaque, making it difficult to interpret such networks or develop trust in their operation. In this article, we introduce SIGN method for modeling the network's hidden layer activity using probabilistic models. The activity patterns in layers of interest are modeled as Gaussian mixture models, and transition probabilities between clusters in consecutive modeled layers are estimated to identify paths of inference. For fully connected networks, the entire layer activity is clustered, and the resulting model is a hidden Markov model. For convolutional layers, spatial columns of activity are clustered, and a maximum likelihood model is developed for mining an explanatory inference graph. The graph describes the hierarchy of activity clusters most relevant for network prediction. We show that such inference graphs are useful for understanding the general inference process of a class, as well as explaining the (correct or incorrect) decisions the network makes about specific images. In addition, SIGN provide interesting observations regarding hidden layer activity in general, including the concentration of memorization in a single middle layer in fully connected networks, and a highly local nature of column activities in the top CNN layers.

Detection of overlapping ultrasonic echoes with deep neural networks.

Ultrasonic Pulse-Echo techniques have a significant role in monitoring the integrity of layered structures and adhesive joints along their service life. However, when acoustically measuring thin layers, the resulting echoes from two successive interfaces overlap in time, limiting the resolution that can be resolved using conventional pulse-echo techniques. Deep convolutional networks have arisen as a promising framework, providing state-of-the-art performance for various signal processing tasks. In this paper, we explore the applicability of deep networks for detection of overlapping ultrasonic echoes. The network is shown to outperform traditional algorithms in simulations for a significant range of echo overlaps, echo pattern variance and noise levels. In addition, experiments on two physical phantoms are conducted, demonstrating superiority of the network over traditional methods for layer thickness estimation.

Leaf Counting: Fusing Network Components for Improved Accuracy.

Leaf counting in potted plants is an important building block for estimating their health status and growth rate and has obtained increasing attention from the visual phenotyping community in recent years. Two novel deep learning approaches for visual leaf counting tasks are proposed, evaluated, and compared in this study. The first method performs counting via direct regression but using multiple image representation resolutions to attend leaves of multiple scales. The leaf count from multiple resolutions is fused using a novel technique to get the final count. The second method is detection with a regression model that counts the leaves after locating leaf center points and aggregating them. The algorithms are evaluated on the Leaf Counting Challenge (LCC) dataset of the Computer Vision Problems in Plant Phenotyping (CVPPP) conference 2017, and a new larger dataset of banana leaves. Experimental results show that both methods outperform previous CVPPP LCC challenge winners, based on the challenge evaluation metrics, and place this study as the state of the art in leaf counting. The detection with regression method is found to be preferable for larger datasets when the center-dot annotation is available, and it also enables leaf center localization with a 0.94 average precision. When such annotations are not available, the multiple scale regression model is a good option.

A unified deep network for beamforming and speckle reduction in plane wave imaging: A simulation study.

Plane Wave Imaging is a fast imaging method used in ultrasound, which allows a high frame rate, but with compromised image quality when a single wave is used. In this work a learning-based approach was used to obtain improved image quality. The entire process of beamforming and speckle reduction was embedded in a single deep convolutional network, and trained with two types of simulated data. The network architecture was designed based on traditional physical considerations of the ultrasonic image formation pipe. As such, it includes beamforming with spatial matched filters, envelope detection, and a speckle reduction stage done in log-signal representation, with all stages containing trainable parameters. The approach was tested on the publicly available PICMUS datasets, achieving axial and lateral full-width-half-maximum (FWHM) resolution values of 0.22 mm and 0.35 mm respectively, and a Contrast to Noise Ratio (CNR) metric of 16.75 on the experimental datasets.

Comparison processes in category learning: from theory to behavior.

Recent studies stressed the importance of comparing exemplars both for improving performance by artificial classifiers as well as for explaining human category-learning strategies. In this report we provide a theoretical analysis for the usability of exemplar comparison for category-learning. We distinguish between two types of comparison -- comparison of exemplars identified to belong to the same category vs. comparison of exemplars identified to belong to two different categories. Our analysis suggests that these two types of comparison differ both qualitatively and quantitatively. In particular, in most everyday life scenarios, comparison of same-class exemplars will be far more informative than comparison of different-class exemplars. We also present behavioral findings suggesting that these properties of the two types of comparison shape the category-learning strategies that people implement. The predisposition for use of one strategy in preference to the other often results in a significant gap between the actual information content provided, and the way this information is eventually employed. These findings may further suggest under which conditions the reported category-learning biases may be overcome.

Spike sorting: Bayesian clustering of non-stationary data.

Spike sorting involves clustering spikes recorded by a micro-electrode according to the source neurons. It is a complicated task, which requires much human labor, in part due to the non-stationary nature of the data. We propose to automate the clustering process in a Bayesian framework, with the source neurons modeled as a non-stationary mixture-of-Gaussians. At a first search stage, the data are divided into short time frames, and candidate descriptions of the data as mixtures-of-Gaussians are computed for each frame separately. At a second stage, transition probabilities between candidate mixtures are computed, and a globally optimal clustering solution is found as the maximum-a-posteriori solution of the resulting probabilistic model. The transition probabilities are computed using local stationarity assumptions, and are based on a Gaussian version of the Jensen-Shannon divergence. We employ synthetically generated spike data to illustrate the method and show that it outperforms other spike sorting methods in a non-stationary scenario. We then use real spike data and find high agreement of the method with expert human sorters in two modes of operation: a fully unsupervised and a semi-supervised mode. Thus, this method differs from other methods in two aspects: its ability to account for non-stationary data, and its close to human performance.