In this work, we design a data synthesis pipeline for airport runway Foreign Object Debris (FOD) detection. Our proposed pipeline constructs a synthetic dataset using diffusion models to generate images based on previously acquired images and newly captured image samples, which commonly exhibit different lighting conditions.We explore various training configurations and conduct a comparative analysis of the results. The findings demonstrate that the proposed pipeline effectively boosts the performance of FOD detection, with the model achieving high performance across precision, recall, and mAP50 metrics.

Coronary artery disease (CAD) is a major cause of mortality, with diagnosis largely dependent on quantitative coronary analysis (QCA) using coronary angiogram (CAG) images. Advances in deep learning have significantly improved medical image processing, creating new possibilities for automated QCA. This paper focuses on applying machine learning techniques to vessel segmentation in CAG images, a crucial first step in QCA. We propose a novel semi-supervised learning approach, introducing a semi-supervised Dice loss with three components: a foreground Dice loss for accurate segmentation, a background Dice loss to reduce false positives, and a semi-supervised Dice loss to generalize to unlabeled areas. We also present a focal masking strategy that adaptively targets underperforming regions, such as minor vessels, ensuring the model focuses on challenging areas to improve overall segmentation performance. To our knowledge, this is the first application of semi-supervised learning for CAG vessel segmentation using deep learning, offering a more efficient and accurate solution that could enhance diagnostic outcomes in clinical settings.

The field of machine learning has been greatly transformed with the advancement of deep artificial neural networks (ANNs) and the increased availability of annotated data. Spiking neural networks (SNNs) have recently emerged as a low-power alternative to ANNs due to their sparsity nature.In this work, we propose a novel hybrid ANN-SNN cotraining framework to improve the performance of converted SNNs. Our approach is a fine-tuning scheme, conducted through an alternating, forward-backward training procedure. We apply our framework to object detection and image segmentation tasks. Experiments demonstrate the effectiveness of our approach in achieving the design goals.

In this paper, we explore the capabilities of a number of deep neural network models in generating whole-brain 3T-like MR images from clinical 1.5T MRIs. The models include a fully convolutional network (FCN) method and three state-of-the-art super-resolution solutions, ESPCN [26], SRGAN [17] and PRSR [7]. The FCN solution, U-Convert-Net, carries out mapping of 1.5T-to-3T slices through a U-Net-like architecture, with 3D neighborhood information integrated through a multi-view ensemble. The pros and cons of the models, as well the associated evaluation metrics, are measured with experiments and discussed in depth. To the best of our knowledge, this study is the first work to evaluate multiple deep learning solutions for whole-brain MRI conversion, as well as the first attempt to utilize FCN/U-Net-like structure for this purpose.

Collision avoidance is a crucial task in vision-guided autonomous navigation. Solutions based on deep reinforcement learning (DRL) has become increasingly popular. In this work, we proposed several novel agent state and reward function designs to tackle two critical issues in DRL-based navigation solutions: 1) smoothness of the trained flight trajectories; and 2) model generalization to handle unseen environments.Formulated under a DRL framework, our model relies on margin reward and smoothness constraints to ensure UAVs fly smoothly while greatly reducing the chance of collision. The proposed smoothness reward minimizes a combination of first-order and second-order derivatives of flight trajectories, which can also drive the points to be evenly distributed, leading to stable flight speed. To enhance the agent’s capability of handling new unseen environments, two practical setups are proposed to improve the invariance of both the state and reward function when deploying in different scenes. Experiments demonstrate the effectiveness of our overall design and individual components.

Speaker separation aims to extract multiple voices from a mixed signal. In this paper, we propose two speaker-aware designs to improve the existing speaker separation solutions. The first model is a speaker conditioning network that integrates speech samples to generate individualized speaker conditions, which then provide informed guidance for a separation module to produce well-separated outputs.The second design aims to reduce non-target voices in the separated speech. To this end, we propose negative distances to penalize the appearance of any non-target voice in the channel outputs, and positive distances to drive the separated voices closer to the clean targets. We explore two different setups, weighted-sum and triplet-like, to integrate these two distances to form a combined auxiliary loss for the separation networks. Experiments conducted on LibriMix demonstrate the effectiveness of our proposed models.

The ultimate goal of speech enhancement is to improve speech quality and intelligibility. Integrating human speech elements into waveform denoising neural networks has proven to be a simple yet effective strategy for this purpose. Such integration, however, has mostly been carried out within supervised learning settings, without taking advantage of the power of the latest self-supervised learning models, which have demonstrated remarkable capability of extracting knowledge from large training sets. In this paper, we present K-SENet, a knowledge-assisted waveform framework for speech enhancement. Wave-U-Net is utilized as the baseline model and the foundation to build our framework. To achieve enhanced intelligibility, we propose a perceptual loss function that relies on self-supervised speech representations pretrained on large datasets, to provide guidance for the baseline network. Wav2vec and PASE are the choices of self-supervised models in this work. Our proposed perceptual loss is calculated upon the perceptual similarities captured by the speech representations. Minimizing this loss would ensure the denoised network outputs sound like clean human speeches. Experiments on the Noisy VCTK and modified TIMIT datasets demonstrate that our K-SENet can significantly improve the perceptual quality of network outputs.

In this paper, we propose a nonlinear distance metric learning scheme based on the fusion of component linear metrics. Instead of merging displacements at each data point, our model calculates the velocities induced by the component transformations, via a geodesic interpolation on a Lie transformation group. Such velocities are later summed up to produce a global transformation that is guaranteed to be diffeomorphic. Consequently, pair-wise distances computed this way conform to a smooth and spatially varying metric, which can greatly benefit k-NN classification. Experiments on synthetic and real datasets demonstrate the effectiveness of our model.

In this paper, we propose a nonlinear metric learning framework to boost the performance of semi-supervised learning (SSL) algorithms. Constructed on top of Laplacian SVM (LapSVM), the proposed method learns a smooth nonlinear feature space transformation that makes the input data points more linearly separable. Coherent point drifting (CPD) is utilized as the geometric model with the consideration of its remarkable expressive power in generating sophisticated yet smooth deformations. Our framework has broad applicability, and it can be integrated with many other SSL classifiers than LapSVM. Experiments performed on synthetic and real world datasets show the effectiveness of our CPD-LapSVM over the state-of-the-art metric learning solutions in SSL.

Shape matching in the spectral domain has gained great popularity in recent years. Most algorithms, however, rely on invariant global spectral embeddings of the shapes to find correspondence, where spatial neighborhood information is not explicitly incorporated into the matching procedure. Misalignments of global as well as local structures are often resulted due to the lack of spatial guidance. In this paper, we identify a number of ambiguities existing in spectral embedding and matching, and subsequently propose a general framework to improve the matching coherence. At the center of the framework is a hybrid spatial-awareness spectral embedding (SASE), which allows various neighborhood and topological information, such as pair-wise distance, relative angles w.r.t. object centers, to be integrated into commute-time (CT) embeddings. A probabilistic expectation maximization (EM) algorithm with imposed regularity is employed to seek an optimal matching of the SASE embeddings. Experimental evaluations of the algorithm on 2D and 3D data demonstrate both the effectiveness and robustness of our approach.

Computing the average anatomy and measuring the anatomical variability within a group of subjects are common practices in Computational Anatomy. In this paper, we propose a statistical analysis framework for 2D/3D shapes. At the core of the framework is a parametric shape representation formulated as a concatenation of skeleton points and the discs centered at the points. This shape representation possesses an excellent capability of capturing both global structures and local details. The constructed Riemannian manifold shape space provides a mathematically sound foundation for various groupwise operations, such as calculating the mean shape and conducting structure-specific normalization. Experiments with 2D shapes and 3D human brain structures show the effectiveness of our framework in calculating the distances among different shapes.

Radial Basis Functions (RBF) are effective in modeling regularization stabilizers, and have been successfully utilized in several point-based registration algorithms. Unfortunately the solutions usually require the inversion of a matrix or solving a linear system, whose computational cost grows rapidly with the increase of the input data size. In this paper, we present a fast and robust approximation remedy for this issue. Our model formulates the registration objective function under the Generalized Radial Basis Function (GRBF) framework w.r.t the cluster centers of one point set. With fewer variables, an computationally efficient registration is achieved, which updates the non-rigid transformation and the correspondence matrix simultaneously. Since the cluster centers often capture the global structure of the point sets very well, enhanced registration robustness is also resulted due to the less likelihood of trapping into local minima. This is especially beneficial in the context of large or/and unevenly distributed data sets. By means of experiments on real and synthetic data, we demonstrate the improvements made over several state-of-the-art solutions.

Comparing different shapes is a fundamental problem in Computational Anatomy (CA), where a rigorous and intrinsic distance metric is key for a shape analysis system to work effectively and consistently. In this paper, we propose a shape comparison and classification framework that consists of two major components. A meridian-based shape representation, stemmed from spectral graph theory, possesses the merit of being able to capture the salient structure property along the direction of maximal shape variations. The meridian extraction algorithm, relying on a discrete approximation of the gradient of the induced Fiedler function, can also be utilized for other purposes, e.g., mesh generation for three-dimensional objects. After projecting the 3D meridians onto a multi-dimensional sphere, similarity/dissimilarity between shapes can be computed based on a Riemannian spherical distance metric. Group statistics, as well as object classification/clustering, can be readily carried out. We demonstrate the effectiveness of our framework with subcortical structures extracted from human brain MR images.

In this paper, we propose a distribution-based active contour model for brain MRI segmentation. As a generalization of the Chan-Vese piecewise-constant model, our solution uses Bayesian a posterior probabilities as the driving forces for curve evolution. Distribution prior, if available, can be seamlessly integrated into the level set evolution procedure. Unlike other region-based active contour models, our solution relaxes the global piecewise-constant assumption, and uses locally varying Gaussians to better account for intensity inhomogeneity and local variations existing in many MR images. More accurate and robust segmentations are therefore achieved. Experiments conducted on synthetic and real brain MRIs demonstrate the improvement made by our model. 1

Active contours or snakes have been extensively utilized in handling image segmentation and classification problems. In traditional active contour models, snake initialization is performed manually by users, and topological changes, such as splitting of the snake, can not be automatically handled. In this paper, we introduce a new method to solve the snake initialization and splitting problem, based on an area segmentation approach: the external force field is segmented first, and then the snake initialization and splitting can be automatically performed by using the segmented external force field. Such initialization and splitting produces multiple snakes, each of which is within the capture range associated to an object and evolved to the object boundary. The external force used in this paper is a gradient vector flow with an edge-preserving property (EPGVF), which can prevent the snakes from passing over weak boundaries. To segment the external force field, we represent it with a graph, and a graph-theory approach can be taken to determine the membership of each pixel. Experimental results establish the effectiveness of the proposed approach.

Mutual information (MI) is currently the most popular match metric in handling the registration problem for multimodality images. However, interpolation artifacts impose deteriorating effects to the accuracy and robustness of MI-based methods. This paper analyzes the generation mechanism of the artifacts inherent in partial volume interpolation (PVI) and shows that the mutual information resulted from PVI is a convex function within each voxel grid. A new joint entropy estimation scheme using prior information is proposed to reduce the artifact effects and we demonstrate the improvements via experiments on misalignments between MR brain scans obtained using different image acquisition protocols.

Automatic registration of multimodal images involves algorithmically estimating the coordinate transformation required to align the data sets. Most existing methods in the literature are unable to cope with registration of image pairs with large nonoverlapping field of view (FOV). We propose a robust algorithm, based on matching dominant local frequency image representations, which can cope with image pairs with large nonoverlapping FOV. The local frequency representation naturally allows for processing the data at different scales/resolutions, a very desirable property from a computational efficiency view point. Our algorithm involves minimizing-over all rigid/affine transformations--the integral of the squared error (ISE or L2 E) between a Gaussian model of the residual and its true density function. The residual here refers to the difference between the local frequency representations of the transformed (by an unknown transformation) source and target data. We present implementation results for image data sets, which are misaligned magnetic resonance (MR) brain scans obtained using different image acquisition protocols as well as misaligned MR-computed tomography scans. We experimently show that our L2E-based scheme yields better accuracy over the normalized mutual information.