Abstract Deep learning-based object detection models have become a preferred choice for crop detection tasks in crop monitoring activities due to their high accuracy and generalization capabilities. However, their high computational demand and large memory footprint pose a challenge for use on mobile embedded devices deployed in crop monitoring settings. Various approaches have been taken to minimize the computational cost and reduce the size of object detection models such as channel and layer pruning, detection head searching, backbone optimization, etc. In this work, we approached computational lightening, model compression, and speed improvement by discarding one or more of the three detection scales of the YOLOv5 object detection model. Thus, we derived up to five separate fast and light models, each with only one or two detection scales. To evaluate the new models for a real crop monitoring use case, the models were deployed on NVIDIA Jetson nano and NVIDIA Jetson Orin devices. The new models achieved up to 21.4% reduction in giga floating-point operations per second (GFLOPS), 31.9% reduction in number of parameters, 30.8% reduction in model size, 28.1% increase in inference speed, with only a small average accuracy drop of 3.6%. These new models are suitable for crop detection tasks since the crops are usually of similar sizes due to the high likelihood of being in the same growth stage, thus, making it sufficient to detect the crops with just one or two detection scales.

Attention Deficit/Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder in childhood that often persists into adulthood. Objectively diagnosing ADHD can be challenging due to the reliance on subjective questionnaires in clinical assessment. Fortunately, recent advancements in artificial intelligence (AI) have shown promise in providing objective diagnoses through the analysis of medical images or activity recordings. These AI-based techniques have demonstrated accurate ADHD diagnosis; however, the growing complexity of deep learning models has introduced a lack of interpretability. These models often function as black boxes, unable to offer meaningful insights into the data patterns that characterize ADHD. OBJECTIVE: This paper proposes a methodology to interpret the output of an AI-based diagnosis system for combined ADHD in age and gender-stratified populations. METHODS: Our system is based on the analysis of 24 hour-long activity records using Convolutional Neural Networks (CNNs) to classify spectrograms of activity windows. These windows are interpreted using occlusion maps to highlight the time-frequency patterns explaining ADHD activity. RESULTS: Significant differences in the frequency patterns between ADHD and controls both in diurnal and nocturnal activity were found for all the populations. Temporal dispersion also presented differences in the male population. CONCLUSION: The proposed interpretation techniques for CNNs highlighted gender- and age-related differences between ADHD patients and controls. Leveraging these differences could potentially lead to improved diagnostic accuracy, especially if a larger and more balanced dataset is utilized. SIGNIFICANCE: Our findings pave the way for the development of an AI-based diagnosis system for ADHD that offers interpretability, thereby providing valuable insights into the underlying etiology of the disease.

This paper proposes a complete convolutional formulation for 2D multimodal pairwise image registration problems based on free-form deformations. We have reformulated in terms of discrete 1D convolutions the evaluation of spatial transformations, the regularization term, and their gradients for three different multimodal registration metrics, namely, normalized cross correlation, mutual information, and normalized mutual information. A sufficient condition on the metric gradient is provided for further extension to other metrics. The proposed approach has been tested, as a proof of concept, on contrast-enhanced first-pass perfusion cardiac magnetic resonance images. Execution times have been compared with the corresponding execution times of the classical tensor product formulation, both on CPU and GPU. The speed-up achieved by using convolutions instead of tensor products depends on the image size and the number of control points considered, the larger those magnitudes, the greater the execution time reduction. Furthermore, the speed-up will be more significant when gradient operations constitute the major bottleneck in the optimization process.

Recently, an efficient implementation of convolution-based free form deformations (FFD) has been proposed for both groupwise 3D monomodal and 2D pairwise multimodal registrations. However, there is still an unmet need in the field for groupwise L-D multimodal registration with L≥2. In this correspondence, we address this need and present a solution for achieving accurate registration using two popular metrics: Renyi entropy and PCA2.

Attention Deficit/Hyperactivity Disorder (ADHD) is the most common neurobehavioral disorder in children and adolescents. However, its etiology is still unknown, and this hinders the existence of reliable, fast and inexpensive standard diagnostic methods. OBJECTIVE: This paper proposes an end-to-end methodology for automatic diagnosis of the combined type of ADHD. METHODS: Diagnosis is based on the analysis of 24 hour-long activity records using Convolutional Neural Networks to classify spectrograms of activity windows. RESULTS: We achieve up to [Formula: see text] average sensitivity, [Formula: see text] specificity and AUC values over [Formula: see text]. Overall, our figures overcome those obtained by actigraphy-based methods reported in the literature as well as others based on more expensive (and not so convenient) acquisition methods. CONCLUSION: These results reinforce the idea that combining deep learning techniques together with actimetry can lead to a robust and efficient system for objective ADHD diagnosis. SIGNIFICANCE: Reliance on simple activity measurements leads to an inexpensive and non-invasive objective diagn-ostic method, which can be easily implemented with daily devices.

The Fast Marching method is widely employed in several fields of image processing. Some years ago a Multi-Stencil version (MSFM) was introduced to improve its accuracy by solving the equation for a set of stencils and choosing the best solution at each considered node. The following work proposes a modified numerical scheme for MSFM to take into account the variation of the local cost, which has proven to be second order. The influence of the stencil set choice on the algorithm outcome with respect to stencil orthogonality and axis swapping is also explored, where stencils are taken from neighborhoods of varying radius. The experimental results show that the proposed schemes improve the accuracy of their original counterparts, and that the use of permutation-invariant stencil sets provides robustness against shifted vector coordinates in the stencil set.

Medical image processing is often limited by the computational cost of the involved algorithms. Whereas dedicated computing devices (GPUs in particular) exist and do provide significant efficiency boosts, they have an extra cost of use in terms of housekeeping tasks (device selection and initialization, data streaming, synchronization with the CPU, and others), which may hinder developers from using them. This paper describes an OpenCL-based framework that is capable of handling dedicated computing devices seamlessly and that allows the developer to concentrate on image processing tasks. The framework handles automatically device discovery and initialization, data transfers to and from the device and the file system and kernel loading and compiling. Data structures need to be defined only once independently of the computing device; code is unique, consequently, for every device, including the host CPU. Pinned memory/buffer mapping is used to achieve maximum performance in data transfers. Code fragments included in the paper show how the computing device is almost immediately and effortlessly available to the users algorithms, so they can focus on productive work. Code required for device selection and initialization, data loading and streaming and kernel compilation is minimal and systematic. Algorithms can be thought of as mathematical operators (called processes), with input, output and parameters, and they may be chained one after another easily and efficiently. Also for efficiency, processes can have their initialization work split from their core workload, so process chains and loops do not incur in performance penalties. Algorithm code is independent of the device type targeted.

Characterization of the noise distribution in magnetic resonance images has multiple applications, including quality assurance and protocol optimization. Noise characterization is particularly important in the presence of parallel imaging acceleration with multi-coil acquisitions, where the noise distribution can contain severe spatial heterogeneities. If the parallel imaging reconstruction is a linear process, an accurate noise analysis can be carried out by taking into account the correlations between all the samples involved. However, for -space-based techniques such as generalized autocalibrating partially parallel acquisition (GRAPPA), the exact analysis has been considered computationally prohibitive due to the very large size of the noise covariance matrices required to characterize the noise propagation from -space to image space. Previously proposed methods avoid this computational burden by formulating the GRAPPA reconstruction as a pixel-wise linear operation performed in the image space. However, these methods are not exact in the presence of non-uniform sampling of -space (e.g., containing a calibration region). For this reason, in this paper, we develop an accurate characterization of the noise distribution for self-calibrated parallel imaging in the presence of arbitrary Cartesian sampling patterns. By exploiting the symmetries and separability in the noise propagation process, the proposed method is computationally efficient and does not require large matrices. Under the assumption of a fixed reconstruction kernel, this method provides the precise distribution of the noise variance for each coil's image. These coil-by-coil noise maps are subsequently combined according to the coil combination approach used in image reconstruction, and therefore can be applied with both complex coil combination and root-sum-of-squares approaches. In this paper, we present the proposed noise characterization method and compare it to previous techniques using Monte Carlo simulations as well as phantom acquisitions.

Ultrasound (US) imaging exhibits considerable difficulties for medical visual inspection and for development of automatic analysis methods due to speckle, which negatively affects the perception of tissue boundaries and the performance of automatic segmentation methods. With the aim of alleviating the effect of speckle, many filtering techniques are usually considered as a preprocessing step prior to automatic analysis methods or visual inspection. Most of the state-of-the-art filters try to reduce the speckle effect without considering its relevance for the characterization of tissue nature. However, the speckle phenomenon is the inherent response of echo signals in tissues and can provide important features for clinical purposes. This loss of information is even magnified due to the iterative process of some speckle filters, e.g., diffusion filters, which tend to produce over-filtering because of the progressive loss of relevant information for diagnostic purposes during the diffusion process. In this paper, we propose an anisotropic diffusion filter with a probabilistic-driven memory mechanism to overcome the over-filtering problem by following a tissue selective philosophy. In particular, we formulate the memory mechanism as a delay differential equation for the diffusion tensor whose behavior depends on the statistics of the tissues, by accelerating the diffusion process in meaningless regions and including the memory effect in regions where relevant details should be preserved. Results both in synthetic and real US images support the inclusion of the probabilistic memory mechanism for maintaining clinical relevant structures, which are removed by the state-of-the-art filters.

We propose a fully 3-D methodology for the computation of myocardial nonviable tissue transmurality in contrast enhanced magnetic resonance images. The outcome is a continuous map defined within the myocardium where not only current state-of-the-art measures of transmurality can be calculated, but also information on the location of nonviable tissue is preserved. The computation is done by means of a partial differential equation framework we have called multi-stencil streamline fast marching. Using it, the myocardial and scarred tissue thickness is simultaneously computed. Experimental results show that the proposed 3-D method allows for the computation of transmurality in myocardial regions where current 2-D methods are not able to as conceived, and it also provides more robust and accurate results in situations where the assumptions on which current 2-D methods are based-i.e., there is a visible endocardial contour and its corresponding epicardial points lie on the same slice-, are not met.

This paper proposes a methodology for the joint alignment of a sequence of images based on a groupwise registration procedure by using a new family of metrics that exploit the expected sparseness of the temporal intensity curves corresponding to the aligned points. Therefore, this methodology is able to tackle the alignment of temporal sequences of images in which the represented phenomenon varies in time. Specifically, we have applied it to the correction of motion in contrast-enhanced first-pass perfusion cardiac magnetic resonance images. The time sequence is elastically registered as a whole by using the aforementioned family of multi-image metrics and jointly optimizing the parameters of the transformations involved. The proposed metrics are able to cope with dynamic changes in the intensity content of corresponding points in the sequence guided by the assumption that these changes allow for a sparse representation in a properly selected frame. Results have shown the statistically significant improvement in the performance of the proposed metric with respect to previous groupwise registration metrics for the problem at hand, which is especially relevant to correct for elastic deformations.

This paper proposes a topology-preserving multiresolution elastic registration method based on a discrete Markov random field of deformations and a block-matching procedure. The method is applied to the object-based interpolation of tomographic slices. For that purpose, the fidelity of a given deformation to the data is established by a block-matching strategy based on intensity- and gradient-related features, the smoothness of the transformation is favored by an appropriate prior on the field, and the deformation is guaranteed to maintain the topology by imposing some hard constraints on the local configurations of the field. The resulting deformation is defined as the maximum a posteriori configuration. Additionally, the relative influence of the fidelity and smoothness terms is weighted by the unsupervised estimation of the field parameters. In order to obtain an unbiased interpolation result, the registration is performed both in the forward and backward directions, and the resulting transformations are combined by using the local information content of the deformation. The method is applied to magnetic resonance and computed tomography acquisitions of the brain and the torso. Quantitative comparisons offer an overall improvement in performance with respect to related works in the literature. Additionally, the application of the interpolation method to cardiac magnetic resonance images has shown that the removal of any of the main components of the algorithm results in a decrease in performance which has proven to be statistically significant.

This paper proposes a novel method to detect anomalies in network traffic, based on a nonrestricted α-stable first-order model and statistical hypothesis testing. To this end, we give statistical evidence that the marginal distribution of real traffic is adequately modeled with α-stable functions and classify traffic patterns by means of a Generalized Likelihood Ratio Test (GLRT). The method automatically chooses traffic windows used as a reference, which the traffic window under test is compared with, with no expert intervention needed to that end. We focus on detecting two anomaly types, namely floods and flash-crowds, which have been frequently studied in the literature. Performance of our detection method has been measured through Receiver Operating Characteristic (ROC) curves and results indicate that our method outperforms the closely-related state-of-the-art contribution described in. All experiments use traffic data collected from two routers at our university-a 25,000 students institution-which provide two different levels of traffic aggregation for our tests (traffic at a particular school and the whole university). In addition, the traffic model is tested with publicly available traffic traces. Due to the complexity of α-stable distributions, care has been taken in designing appropriate numerical algorithms to deal with the model.

In this paper, we propose a new method for collision handling between 3-D deformable (organs) and rigid (surgical tools) objects valid for nonstructured interaction scenes and, specifically, for laparoscopic surgery simulators. During the simulation step, vertices of the organ detected as collided must be accurately shifted out of the tool to elude a visual interpenetration. The proposed approach obtains the new position of each collided vertex of the organ taking into account both kinematic information of the surgical tool and geometric information of the organ surface that surrounds the vertex under analysis. Three parameters inferred from a fuzzy feedback system weigh the nature of the tool motion with respect to the organ. Experimental results show that the solution proposed in this paper is able to avoid the interpenetration among the multiple colliding points efficiently with physically and spatially coherent results.

A matrix procedure to obtain a hierarchical decomposition of a multiple-input–single-output (MISO) fuzzy system is proposed. It is based on the fast inference using transition matrices (FITM) framework recently described. The resulting hierarchical system is totally equivalent to the original one in terms of input–output relation. A study of the rule-reduction capability of this system is carried out. Additionally, this paper describes a rule-base definition procedure of each of the systems resulting within the hierarchy. This procedure is based on closed-form analytical expressions and it does not need any manual interaction.

A new method for noise filtering in images that follow a Rician model-with particular attention to magnetic resonance imaging-is proposed. To that end, we have derived a (novel) closed-form solution of the linear minimum mean square error (LMMSE) estimator for this distribution. Additionally, a set of methods that automatically estimate the noise power are developed. These methods use information of the sample distribution of local statistics of the image, such as the local variance, the local mean, and the local mean square value. Accordingly, the dynamic estimation of noise leads to a recursive version of the LMMSE, which shows a good performance in both noise cleaning and feature preservation. This paper also includes the derivation of the probability density function of several local sample statistics for the Rayleigh and Rician model, upon which the estimators are built.

A method to identify different tissues in ultrasound images is presented. Assuming a Rayleigh model for speckle, the different tissues present in the image will be related to the Rayleigh sigma parameter, with different values for each different tissue. The parameter is locally estimated using well known estimation methods for the Rayleigh distribution, such as the maximum likelihood estimator. A significant increase in the separability of the different tissues present in the image is then achieved. A subsequent clustering process based on the position of the maxima of the estimator distribution allows the classification of each pixel in a tissue class.

In this paper, we focus on the problem of speckle removal by means of anisotropic diffusion and, specifically, on the importance of the correct estimation of the statistics involved. First, we derive an anisotropic diffusion filter that does not depend on a linear approximation of the speckle model assumed, which is the case of a previously reported filter, namely, SRAD. Then, we focus on the problem of estimation of the coefficient of variation of both signal and noise and of noise itself. Our experiments indicate that neighborhoods used for parameter estimation do not need to coincide with those used in the diffusion equations. Then, we show that, as long as the estimates are good enough, the filter proposed here and the SRAD perform fairly closely, a fact that emphasizes the importance of the correct estimation of the coefficients of variation.

Fast inference using transition matrices (FITMs) was recently presented as a new fast algorithm for performing inferences in fuzzy systems. However, this algorithm only applies to the standard additive model (SAM) using the sum-product inference composition. In this paper, we show how this methodology carries over to more general Mamdani fuzzy systems, i.e., those using t-norms and t-conorms that satisfy a requirement of distributivity. In addition, FITM, as originally described, requires inputs to be fuzzy sets. We here extend this methodology to handle crisp inputs as well.

A matrix inference method for fuzzy systems (FITM) is used to deal with hierarchical fuzzy systems (HFSs). The contribution of this paper is twofold: on one hand, we show that a HFS (either serial or parallel) can be easily converted into a single fuzzy system, and implemented via FITM. On the other, we show that a serial HFS can be obtained out of a complicated multiple-input single-output fuzzy system. An example of such a conversion for a three-input system is included.

Fast inference using transition matrices (FITM) is a new fast algorithm for performing inferences in fuzzy systems. It is based on the assumption that fuzzy inputs can be expressed as a linear composition of the fuzzy sets used in the rule base. This representation let us interpret a fuzzy set as a vector, so we can just work with the coordinates of it instead of working with the whole set. The inference is made using transition matrices. The key of the method is the fact that a lot of operations can be precomputed offline to obtain the transition matrices, so actual inferences are reduced to a few online matrix additions and multiplications. The algorithm is designed for the standard additive model using the sum-product inference composition.

In this paper we analyze a result previously published about a comparison between two statistical tests used for evaluation of boundary detection algorithms on medical images. We conclude that the statement made by Chalana and Kim (1997) about the performance of the percentage test has a weak theoretical foundation, and according to our results, is not correct. In addition, we propose a one-sided hypothesis test for which the acceptance region can be determined in advance, as opposed to the two-sided confidence intervals proposed in the original paper, which change according to the estimated quantity.

This paper addresses the problem of finding matching points in stereo image pairs, i.e., the problem of correspondence. Even though this topic is well-known, a complete probabilistic formulation of it using psychovisual cues is still missing. We propose a novel Bayesian model based on Markov Random Fields (MRFs); the prior energy function is built in terms of the probability density function (pdf) of the disparity gradient. This pdf has never been reported in the past. The likelihood energy function is defined in terms of the pdf of the square normalized cross covariance between any two matching points. The stereo correspondence map is then obtained as the MAP estimator of the posterior field. Comparative results with methods previously reported, show the adequacy of the general model here proposed, and a good compromise between deterministic and stochastic images is attained.

In this paper we propose an alternative implementation of the concept of fuzzy granule. Granules are defined in terms of the degree of overlap with other granules, as opposed to by assigning (somehow arbitrarily) membership values to each and every point of the university of discourse in which the granule is defined. We believe this alternative definition is much closer to the human way of thinking. Two examples of real world applications illustrate this new definition.

This paper presents an algorithm for automatically detecting bone contours from hand radiographs using active contours. Prior knowledge is first used to locate initial contours for the snakes inside each bone of interest. Next, an adaptive snake algorithm is applied so that parameters are properly adjusted for each bone specifically. We introduce a novel truncation technique to prevent the external forces of the snake from pulling the contour outside the bones boundaries, yielding excellent results.

We carry out a detailed analysis of a fuzzy version of Reid's classical multiple hypothesis tracking (MHT) algorithm. Our fuzzy version is based on well-known fuzzy feedback systems, but the fact that the system we describe is specialized for likelihood discrimination makes this study particularly novel. We discuss several techniques for rule activation. One of them, namely the sum-product, seems particularly useful for likelihood management and its linearity makes it tractable for further analysis. Our analysis is performed in two stages: 1) we demonstrate that, with appropriately chosen rules, our system can discriminate the correct hypothesis; and 2) the steady-state behavior with a constant input is characterized analytically. This enables us to establish the optimality of the sum-product method and it also gives a simple procedure to predict the system's behavior as a function of the rule base. We believe this fact can be used to devise a simple procedure for fine-tuning the rule base according to the system designer's needs. The application driving our fuzzy MHT implementation and analysis is the tracking of natural language text-based messages. This application is used as an example throughout the paper.

An anisotropic diffusion filter controlled by fuzzy rules is presented. The proposed filter is based in the Perona-Malik technique, using fuzzy reasoning to calculate the diffusion coefficient which controls the whole diffusion. The method has the advantage that it can be used for both smoothing and noise cleaning, as well as edge enhancement. This new approach also allows us to model the diffusion process through a rule base to have a better performance. Some examples are given to illustrate the effectiveness of the proposed technique.

In this paper we address the problem of ultrasound volume data segmentation by means of active contours. The segmentation is performed on an slice-by-slice basis but exploiting the high correlation that exists across nearby slices. We have used active rays to alleviate the computational load of the optimization process. Classical energy functions have trouble finding the optimal solution in speckle imagery. We therefore propose new energy functions that work well under this type of images. Moreover, we propose both a probabilistic relaxation scheme and a heuristic search of the minimal energy configuration to end up making a hybrid use of both. We compare two rendered surfaces to illustrate the smoothness of our reconstruction.

In this paper we present a novel framework for posture and gesture recognition performed by human articulated limbs, focusing the exposition on a notable case study: the human hand. The method relies on visual cues located on the surface of the body which are to be 3D reconstructed by means of stereo. A simple hand geometrical model and a visibility ordering scheme provide the regularization information needed to solve the posture recognition problem (the instantaneous hand configuration) in the presence of occlusions. These recognized postures feed a gesture (sequence of postures with semantic content) recognition system based on a clustering of postures paths and on a Viterbi decoder which chooses the one with the maximum likelihood. Some results are presented for a simple vocabulary consisting of sequences of signs taken from the Spanish hearing-impaired alphabet.

We address the problem of stereo matching in the context of Markov random fields (MRFs) and the Bayesian approach. Although MRFs are used as a tool intended to formulate the problem in a way that allows the application of a powerful optimization technique called simulated annealing, the bases for the formulation of the stereo matching problem from a statistical and Markovian point of view are also stated.

In this paper we show the benefits of applying hypothesis testing to the problem of texture segmentation. In our approach, hypothesis testing is used at two different stages that help to reduce the computational burden associated to iterative methods commonly used in image processing. Specifically, hypothesis testing is used to initially estimate the number of regions the image must be divided into, and to determine a set of points that will remain unchanged after the Markovian postprocessing scheme. These fixed points will contribute to reduce the number of iterations required by the Markovian stage and introduce geometry constraints that will reduce the boundary distortion caused by the stochastic procedure.