Efficient deep learning systems are increasingly essential for automated quality inspection in the agri-food industry. This work systematically investigates the impact of optimized data-loading and augmentation pipelines on both training efficiency and predictive performance of convolutional neural networks for fruit defect classification. Benchmark experiments compare CPU-based preprocessing, multithreaded tf.data pipelines, GPU-accelerated workflows, and NVIDIA DALI, showing up to a 16× reduction in training time together with significantly improved GPU utilization. Building on these findings, the optimized pipeline is deployed on a large-scale industrial dataset comprising more than 3 million tangerine image patches. Carefully designed augmentation strategies—including geometric transformations and color perturbations—are introduced to enhance data diversity while preserving the intrinsic visual characteristics of the product. The substantial reduction in training time enables a more efficient exploration of candidate architectures through a tailored Neural Architecture Search (NAS) framework designed for resource-constrained industrial settings. The proposed framework explores internal CNN hyperparameters while preserving architectural depth to satisfy real-time inference constraints. To reduce the computational cost of NAS, a Random Forest–based performance predictor is trained on early-epoch indicators such as the F1-score and used to rapidly screen candidate models. A genetic algorithm is then employed to efficiently explore the search space and identify high-performing configurations. Experimental results demonstrate that the proposed end-to-end workflow significantly accelerates the model development cycle while maintaining or modestly improving classification accuracy. While the reductions in training time are substantial, the predictive-performance improvements observed through NAS are comparatively modest and should be interpreted primarily as evidence that the proposed framework can identify competitive configurations under industrial deployment constraints. The resulting framework provides a practical and scalable workflow for developing and deploying automated visual inspection systems in industrial agri-food production lines.

In this work, we explore a novel edge detection method based on a Quantum Fuzzy Inference Engine implemented on a real quantum computer. We address an industrial defect detection problem using a proprietary fruit dataset. Our approach combines a hybrid quantum-classical lookup table for edge detection with a classical contour tracking algorithm and is compared against state-of-the-art classical methods. Our hybrid model achieves comparable precision to its classical counterpart while requiring fewer operations. Moreover, comparisons with other leading classical techniques confirm the reliability of our approach, as it delivers very similar results.

Brain cells are arranged in laminar, nuclear, or columnar structures, spanning a range of scales. Here, we construct a reliable cell census in the frontal lobe of human cerebral cortex at micrometer resolution in a magnetic resonance imaging (MRI)–referenced system using innovative imaging and analysis methodologies. MRI establishes a macroscopic reference coordinate system of laminar and cytoarchitectural boundaries. Cell counting is obtained with a digital stereological approach on the 3D reconstruction at cellular resolution from a custom-made inverted confocal light-sheet fluorescence microscope (LSFM). Mesoscale optical coherence tomography enables the registration of the distorted histological cell typing obtained with LSFM to the MRI-based atlas coordinate system. The outcome is an integrated high-resolution cellular census of Broca’s area in a human postmortem specimen, within a whole-brain reference space atlas.

Semantic segmentation of neuronal structures in 3D high-resolution fluorescence microscopy imaging of the human brain cortex can take advantage of bidimensional CNNs, which yield good results in neuron localization but lead to inaccurate surface reconstruction. 3D CNNs on the other hand would require manually annotated volumetric data on a large scale, and hence considerable human effort. Semi-supervised alternative strategies which make use only of sparse annotations suffer from longer training times and achieved models tend to have increased capacity compared to 2D CNNs, needing more ground truth data to attain similar results. To overcome these issues we propose a two-phase strategy for training native 3D CNN models on sparse 2D annotations where missing labels are inferred by a 2D CNN model and combined with manual annotations in a weighted manner during loss calculation.

Although neuronal density analysis on human brain slices is available from stereological studies, data on the spatial distribution of neurons in 3D are still missing. Since the neuronal organization is very inhomogeneous in the cerebral cortex, it is critical to map all neurons in a given volume rather than relying on sparse sampling methods. To achieve this goal, we implement a new tissue transformation protocol to clear and label human brain tissues and we exploit the high-resolution optical sectioning of two-photon fluorescence microscopy to perform 3D mesoscopic reconstruction. We perform neuronal mapping of 100mm3 human brain samples and evaluate the volume and density distribution of neurons from various areas of the cortex originating from different subjects (young, adult, and elderly, both healthy and pathological). The quantitative evaluation of the density in combination with the mean volume of the thousands of neurons identified within the specimens, allow us to determine the layer-specific organization of the cerebral architecture.

Cells are not uniformly distributed in the human cerebral cortex. Rather, they are arranged in a regional and laminar fashion that span a range of scales. Here we demonstrate an innovative imaging and analysis pipeline to construct a reliable cell census across the human cerebral cortex. Magnetic resonance imaging (MRI) is used to establish a macroscopic reference coordinate system of laminar and cytoarchitectural boundaries. Cell counting is obtained with both traditional immunohistochemistry, to provide a stereological gold-standard, and with a custom-made inverted confocal light-sheet fluorescence microscope (LSFM) for 3D imaging at cellular resolution. Finally, mesoscale optical coherence tomography (OCT) enables the registration of the distorted histological cell typing obtained with LSFM to the MRI-based atlas coordinate system.

Example dataset containing light-sheet selective plane illumination microscopy (SPIM) data to illustrate BIDS convention.BIDS version 1.6.0 - Microscopy BEP031 version 0.0.5 (2021-09-07T14:20:00)

The 3D analysis of the human brain architecture at cellular resolution is still a big challenge. In this work, we propose a pipeline that solves the problem of performing neuronal mapping in large human brain samples at micrometer resolution. First, we introduce the SWITCH/TDE protocol: a robust methodology to clear and label human brain tissue. Then, we implement the 2.5D method based on a Convolutional Neural Network, to automatically detect and segment all neurons. Our method proved to be highly versatile and was applied successfully on specimens from different areas of the cortex originating from different subjects (young, adult and elderly, both healthy and pathological). We quantitatively evaluate the density and, more importantly, the mean volume of the thousands of neurons identified within the specimens. In conclusion, our pipeline makes it possible to study the structural organization of the brain and expands the histopathological studies to the third dimension.

We still lack a detailed map of the anatomical disposition of neurons in the human brain. A complete map would be an important step for deeply understanding the brain function, providing anatomical information useful to decipher the neuronal pattern in healthy and diseased conditions. Here, we present several important advances towards this goal, obtained by combining a new clearing method, advanced Light Sheet Microscopy and automated machine-learning based image analysis. We perform volumetric imaging of large sequentially stained human brain slices, labelled for two different neuronal markers NeuN and GAD67, discriminating the inhibitory population and reconstructing the brain connectivity.

Using a custom light sheet fluorescence microscope, we image large stained human brain portions, labelled for NeuN and GAD67 neuronal markers, discerning the inhibitory population via neural-network based image analysis and exposing the brain connectivity.

The three-dimensional reconstruction of large volumes of the human neural networks at cellular resolution is one of the biggest challenges of our days. Commonly, fine slices of samples marked with colorimetric techniques are individually imaged. This approach in addition to being time-consuming does not consider space cell organization, leading to loss of information. The aim of this work was to develop a methodology that allows analyzing the cytoarchitecture of the human brain in three dimensions at high resolution. In particular, we exploit the possibility of combining high-resolution 3D imaging techniques with clearing methodologies. We successfully integrate the SWITCH immunohistochemistry technique with the TDE clearing method to image a large volume of human brain tissue with two-photon fluorescence microscopy. In conclusion, this new approach enables to characterize large human brain specimens with high-resolution optical techniques, giving the possibility to expand the histological studies to the third dimension.

Deciphering brain architecture at a system level requires the ability to quantitatively map its structure with cellular and subcellular resolution. Besides posing significant challenges to current optical microscopy methods, this ambitious goal requires the development of a new generation of tools to make sense of the huge number of raw images generated, which can easily exceed several TeraBytes for a single sample. We present an integrated pipeline enabling transformation of the acquired dataset from a collection of voxel gray levels to a semantic representation of the sample. This pipeline starts with a software for image stitching that computes global optimal alignment of the 3D tiles. The fused volume is then accessed virtually by means of a dedicated API (Application Programming Interface). The virtually fused volume is then processed to extract meaningful information. We demonstrate two complementary approaches based on deep convolutional networks. In one case, a 3D conv-net is used to 'semantically deconvolve' the image, allowing accurate localization of neuronal bodies with standard clustering algorithms (e.g. mean shift). The scalability of this approach is demonstrated by mapping the spatial distribution of different neuronal populations in a whole mouse brain with single-cell resolution. To go beyond simple localization, we exploited a 2D conv-net estimating for each pixel the probability of being part of a neuron. The output of the net is then processed with a contour finding algorithm, obtaining reliable segmentation of cell morphology. This information can be used to classify neurons, expanding the potential of chemical labeling strategies.

We present a software pipeline for high throughput stitching and processing of high-resolution tomographies of whole mouse brains. We then employ machine learning techniques for automatic segmentation and classification of neurons in the acquired datasets.

Quantitative analysis of brain cytoarchitecture requires effective and efficient segmentation of the raw images. This task is highly demanding from an algorithmic point of view, because of the inherent variations of contrast and intensity in the different areas of the specimen, and of the very large size of the datasets to be processed. Here, we report a machine vision approach based on Convolutional Neural Networks (CNN) for the near real-time segmentation of neurons in three-dimensional images with high specificity and sensitivity. This instrument, together with high-throughput sample preparation and imaging, can lay the basis for a quantitative revolution in neuroanatomical studies.

Non Destructive Testing (NDT) is one of the most important aspect in modern manufacturing companies. Automation of this task improves productivity and reliability of distribution chains. We present an optimized implementation of common pattern recognition algorithms that performs NDT on factory products. To the aim of enhancing the industrial integration, our implementation is highly optimized to work on SoC-based (System on Chip: an integrated circuit that integrates all components of a computer into a single chip.) hardware and we worked with the initial idea of an overall design for these devices. While perfectly working on general purpose SoCs, the best performances are achieved on GPU accelerated ones. We reached the notable performance of a PC-based workstation by exploiting technologies like CUDA and BLAS for embedded SoCs. The test case is a collection of toy scenarios commonly found in manufacturing companies.

Precision viticulture has been developing very rapidly in the last two decades. Operational systems commonly use GIS, spatial interpolation, modelling, real time geolocalization, remote sensing and more recently, drone technology. Whereas precision viticulture is all about high space and time resolution of either measurements, analysis, diagnostics or decision support, few operational systems are built in an integrative way, combining many techniques and data sources. Within the European VINTAGE project, hourly time step climate data measurements, topography-based spatial interpolation, mechanistic soil, plant and disease models, remote sensing information and a decision support system are integrated within a GIS server to provide an operational tool for sustainable vineyard management for the grapegrower. The current study presents the overall system framework and first examples of model validation

Peer-to-peer (P2P) computing already accounts for a large part of the traffic on the Internet, and it is likely to become as ubiquitous as current client/server architectures in next generation information systems. This paper addresses a central problem of P2P systems: the design of an optimal overlay communication network for a set of processes on the Internet. Such a network defines membership to the P2P group and allows for members to disseminate information within the group. The problem, named the membership overlay problem (MOP), can be formulated as a dynamic optimization problem where classical combinatorial optimization techniques must face the further challenge of time-varying input data. This paper proposes an innovative, fully distributed, and asynchronous subgradient optimization algorithm for the Lagrangean relaxation of the MOP, which can run online in fully decentralized P2P systems, and integrates it with a distributed heuristic that can achieve sound hot-start states for fast response to varying network structures.

Decomposition techniques are well-known as a means for obtaining tight lower bounds for combinatorial optimization problems, and thus as a component for solution methods. Moreover a long-established research literature uses them for defining problem-specific heuristics. More recently it has been observed that they can be the basis also for designing metaheuristics. This tutorial elaborates this last point, showing how the three main decomposition techniques, namely Dantzig-Wolfe, Lagrangean and Benders decompositions, can be turned into model-based, dual-aware metaheuristics. A well known combinatorial optimization problem, the Single Source Capacitated Facility Location Problem, is then chosen for validation, and the implemented codes of the proposed algorithms are benchmarked on standard instances from literature.

Matheuristics are heuristic algorithms made by the interoperation of metaheuristics and mathematic programming (MP) techniques. An essential feature is the exploitation in some part of the algorithms of features derived from the mathematical model of the problems of interest, thus the definition "model-based metaheuristics" appearing in the title of some events of the conference series dedicated to matheuristics [1]. The topic has attracted the interest of a community of researchers, and this led to the publication of dedicated volumes and journal special issues, [13], [14], besides to dedicated tracks and sessions on wider scope conferences. The increasing maturity of the area permits to outline some trends and possibilities offered by matheuristic approaches. A word of caution is needed before delving into the subject, because obviously the use of MP for solving optimization problems, albeit in a heuristic way, is much older and much more widespread than matheuristics. However, this is not the case for metaheuristics, and also the very idea of designing MP methods specifically for heuristic solution has innovative traits, when opposed to exact methods which turn into heuristics when enough computational resources are not available.