editorial reviewed

Longitudinal datasets like the Parkinson’s Progression Markers Initiative (PPMI) face critical challenges from missing data and privacy constraints. This paper introduces PPMI-Benchmark, the first comprehensive framework evaluating 12 imputation methods and 6 synthetic data generation techniques across clinical, demographic, and biomarker variables in Parkinson’s disease research. We implement advanced methods including HyperImpute (ensemble optimization), VaDER (variational deep embedding), and conditional tabular GANs (CTGAN), evaluating them through novel metrics integrating sliced Wasserstein distance (dSW = 0.039±0.012), temporal consistency analysis, and clinical validity constraints. Our results demonstrate HyperImpute’s superiority in imputation accuracy (MAE=5.16 vs. 5.19–5.57 for baselines), while CTGAN achieves optimal distribution fidelity (SWD=0.039 vs. 0.062–0.146). Crucially, we reveal persistent demographic biases in cognitive scores, with age-related imputation errors increasing by 23% for patients over 70, and propose mitigation strategies. The framework provides actionable guidelines for selecting data completion strategies based on missingness patterns (MCAR/MAR/MNAR), computational constraints, and clinical objectives, advancing reproducibility and fairness in neurodegenerative disease research. Validated on 1,483 PPMI participants, our work addresses emerging needs in healthcare AI governance and synthetic data interoperability for multi-center collaborations.

The increasing demand for data in machine learning raises significant privacy concerns.Federated Learning (FL) enables multiple entities to train models collaboratively without sharing raw data.However, centralized FL (CFL) relies on a central server, making it vulnerable to poisoning attacks and single points of failure (SPOF).Decentralized FL (DFL) addresses these issues by removing the central server.This paper proposes a novel DFL architecture integrating blockchain for resisting attacks and Multi-Party Computation (MPC) for secure model parameter transfer.This architecture enhances security and confidentiality in collaborative learning without compromising result quality.

peer reviewed

Abstract Deep learning presents an efficient set of methods that allow learning from massive volumes of data using complex deep neural networks. To facilitate the design and implementation of algorithms, deep learning frameworks provide a high‐level programming interface. Based on these frameworks, new models, and applications are able to make better and better predictions. One type of deep learning application is the Internet of Things that can gather a continuous flow of data, which causes an explosion of the amount of data. Therefore, to handle this data management issue, computation technologies can offer new perspectives to analyze more data with more complex models. In this context, a cluster of computers can operate to quickly deliver a model or to enable the design of a complex neural network spread among computers. An alternative is to distribute a deep learning task with HPC cloud computing resources and to scale cluster in order to quickly and efficiently train a neural network. As a first step to design an infrastructure aware framework which is able to scale the computing nodes, this work aims to review and analyze the state‐of‐the‐art frameworks by collecting device utilization data during the training task. We gather information about the CPU, RAM and the GPU utilization on deep learning algorithms with and without multi‐threading. The behavior of each framework is discussed and analyzed in order to shed light on the strengths and weaknesses of the different deep learning frameworks.

Sketch-based image retrieval (SBIR) solutions are attracting increased interest in the field of computer vision. These solutions provide an intuitive and powerful tool to retrieve images in large-scale image databases. In this paper, we conduct a comprehensive study of classic triplet CNN training pipelines within the SBIR context. We study the impact of embeddings normalization, model sharing, margin selection, batch size, hard mining selection and the evolution of the number of hard triplets during training to propose several avenues for improvement. We also propose dropout column, an adaptation of dropout for triplet network and similar pipelines. In addition, we also introduce a novel approach to build state-of-the-art SBIR solutions that can be used with low power systems. The whole study is conducted using The Sketchy Database, a large-scale SBIR database. We carry out a series of experiments and show that adopting a few simple modifications enhances significantly existing SBIR pipelines (faster training & higher accuracy). Our study enables us to propose an enhanced pipeline that outperforms previous state-of-the-art on the Sketchy Database by a significant margin (a recall of 53.92% compared to 46.2% at k = 1) and reaches almost human performance (54.27%) on a large-scale benchmark.

Today, agricultural practices are supported by bio-informatics and emerging technologies such as remote sensing, cloud computing and the Internet of Things (IoT), which leads to the concept of “Smart Farming”. Smart farming is a cycle of intelligent detection and monitoring, analysis and planning, as well as control of agricultural operations using a cloud-based event management system. In this paper, we propose WALLeSMART, a cloud-based framework built to capitalize the efforts invested in building smart farming management systems, applied to the Wallonia region of Belgium. The framework proposes an architecture to address the challenges of acquisition, processing, and visualization of massive amounts of data, in both batch and real-time basis. An initial prototype has been developed and tested with various farms and shows prominent results.

Summary Digital phenotyping is an emergent science mainly based on imagery techniques. The tremendous amount of data generated needs important cloud computing for their processing. The coupling of recent advance of distributed databases and cloud computing offers new possibilities of big data management and data sharing for the scientific research. In this paper, we present a solution combining a lambda architecture built around Apache Druid and a hosting platform leaning on Apache Mesos. Lambda architecture has already proved its performance and robustness. However, the capacity of ingesting and requesting of the database is essential and can constitute a bottleneck for the architecture, in particular, for in terms of availability and response time of data. We focused our experimentation on the response time of different databases to choose the most adapted for our phenotyping architecture. Apache Druid has shown its ability to respond to typical queries of phenotyping applications in times generally inferior to the second.

Summary The last few years have been marked by the presence of very large sets of images and videos in our everyday lives. These multimedia objects have a very fast frequency of creation and sharing since images and videos can come from different devices such as smartphones, satellites, cameras, or drones. They are generally used to illustrate objects in different situations (public areas, train stations, hospitals, political and sport events and competitions, etc). As consequence, image and video processing algorithms have got increasing importance for several computer vision applications that should be adapted for managing large‐scale volumes and exploiting high performance computing resources (local or cloud). In this work, we propose a cloud‐based toolbox (platform) for computer vision applications. This platform integrates a toolbox of image and video processing algorithms that can (i) exploit high performance computing cloud resources, (ii) execute applications in real time, and (iii) manage large‐scale database using Big Data technologies. The related libraries and hardware drivers are automatically integrated and configured in order to offer to users an access to the different applications without the need to download, install, and configure software or hardware. Experiments were conducted using three kinds of applications: (i) image and video processing applications, (ii) deep learning techniques for images classification and multiobject localization, and (iii) images indexation and retrieval. These experiments demonstrated the interest of our platform for sharing, in an efficient way, our scientific contributions and annotated databases in order to improve the quality and performance of computer vision applications.

This paper presents a new method for learning the structure of Bayesian Networks. Broadly speaking, we leverage the Branch and Bound (B&B) to derive the best Directed Acyclic Graphs (DAGs) that describes the structure of the network. Our contribution consists in introducing two main heuristics: the first one allows the selection of the graph that has the best score among those that contain less cycles, the second one eliminates the shortest cycle from the selected graph; it aims to reduce the number of explored nodes. Our experimental study asserts that the suggested proposal improves the results for multiple data sets. These facts are confirmed by the reduction of the computation time and the memory overhead.

Summary Nowadays, image and video processing applications have become widely used in many domains related to computer vision. Indeed, they can come from cameras, smartphones, social networks, or from medical devices. Generally, these images and videos are used for illustrating people or objects (cars, trains, planes, etc) in many situations such as airports, train stations, public areas, sport events, and hospitals. Thus, image and video processing algorithms have got increasing importance, they are required from various computer vision applications such as motion tracking, real time event detection, database (images and videos) indexation, and medical computer‐aided diagnosis methods. The main inconvenient of image and video processing applications is the high intensity of computation and the complex configuration and installation of the related materials and libraries. In this paper, we propose a new framework that allows users to select in a smart and efficient way the computing units (CPU or/and GPU) in a cloud‐based platform, in case of processing one image (or one video in real time) or many images (or videos). This framework enables to affect the local or remote computing units for calculation after analyzing the type of media and the algorithm complexity. The framework disposes of a set of selected CPU and GPU‐based computer vision methods, such as image denoising, histogram computation, features descriptors (SIFT, SURF), points of interest extraction, edges detection, silhouette extraction, and sparse and dense optical flow estimation. These primitive functions are exploited in various applications such as medical image segmentation, videos indexation, real time motion analysis, and left ventricle segmentation and tracking from 2D echocardiography. Experimental results showed a global speedup ranging from 5× to 273×(compared to CPU versions) as result of the application of our framework for the above‐mentioned methods. In addition to these performances, the parallel and heterogeneous implementations offered lower power consumption as result of the fast treatment.

Freehand sketches are a simple and powerful tool for communication. They are easily recognized across cultures and suitable for various applications. In this paper, we use deep convolutional neural networks (ConvNets) to address sketch-based image retrieval (SBIR). We first train our ConvNets on sketch and image object recognition in a large scale benchmark for SBIR (the sketchy database). We then conduct a comprehensive study of ConvNets features for SBIR, using a kNN similarity search paradigm in the ConvNet feature space. In contrast to recent SBIR works, we propose a new architecture the quadruplet networks which enhance ConvNet features for SBIR. This new architecture enables ConvNets to extract more robust global and local features. We evaluate our approach on three large scale datasets. Our quadruplet networks outperform previous state-of-the-art on two of them by a significant margin and gives competitive results on the third. Our system achieves a recall of 42.16% (at k=1) for the sketchy database (more than 5% improvement), a Kendal score of 43.28Τb on the TU-Berlin SBIR benchmark (close to 6Τb improvement) and a mean average precision (MAP) of 32.16% on Flickr15k (a category level SBIR benchmark).

peer reviewed

Freehand sketches are an interesting universal form of visual representation. Sketching has become easily accessible with many of the devices that we use on a daily basis. In this paper, we propose a system for real-time sketch recognition and similarity search. Our system is able to recognize partial sketches from 250 object categories. It is then able to retrieve similar sketches but also images/photographs. In this work, we propose to use deep convolutional neural networks (ConvNets) for partial sketch recognition and feature extraction. Features are extracted from sketches and image contours in order to be used as a basis for similarity search using k-Nearest Neighbors (kNN). Our system demonstrates promising results in identifying similar images, and could be integrated in larger content-based search engines.

Knowledge of left atrial (LA) anatomy is important for atrial fibrillation ablation guidance, fibrosis quantification and biophysical modelling. Segmentation of the LA from Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) images is a complex problem. This manuscript presents a benchmark to evaluate algorithms that address LA segmentation. The datasets, ground truth and evaluation code have been made publicly available through the http://www.cardiacatlas.org website. This manuscript also reports the results of the Left Atrial Segmentation Challenge (LASC) carried out at the STACOM'13 workshop, in conjunction with MICCAI'13. Thirty CT and 30 MRI datasets were provided to participants for segmentation. Each participant segmented the LA including a short part of the LA appendage trunk and proximal sections of the pulmonary veins (PVs). We present results for nine algorithms for CT and eight algorithms for MRI. Results showed that methodologies combining statistical models with region growing approaches were the most appropriate to handle the proposed task. The ground truth and automatic segmentations were standardised to reduce the influence of inconsistently defined regions (e.g., mitral plane, PVs end points, LA appendage). This standardisation framework, which is a contribution of this work, can be used to label and further analyse anatomical regions of the LA. By performing the standardisation directly on the left atrial surface, we can process multiple input data, including meshes exported from different electroanatomical mapping systems.

This paper describes a new method of segmentation and identification of individual vertebrae in medical images. The final goal of the application is to determine vertebrae motion induced by their movement between two or several positions. For that, X-ray images of the spinal columns are analysed in order to extract vertebrae contours. We present a new image segmentation approach based on a preliminary selection of vertebrae regions. We use these regions information to identify each individual vertebra by its contour. For the edge detection task, we propose a polar signature representation of the contour using the image gradient of each region. After this, we apply an edge closing method exploiting polynomial fitting.