Large Vessel Occlusion is one of the most critical neurological emergencies in stroke care, requiring rapid and accurate diagnosis to optimize clinical outcomes. Automated detection tools have demonstrated the potential to significantly reduce treatment time, thereby improving patient prognosis. In this paper, we propose a novel two-stage deep learning approach for detecting large vessel occlusion and assessing its volume directly from computed tomography angiography. The first stage employs a two-dimensional convolutional neural network-based detector, built upon GravityNet, specifically adapted for single lesion detection with a novel pixel-based configuration. The second stage applies a three-dimensional false-positive reduction technique to refine predictions within the brain volume. Our method achieves 80% sensitivity at two false positives per case, demonstrating its robustness and effectiveness in detecting large vessel occlusions on computed tomography angiography.

The widespread adoption of electronic health records (EHRs) offers a valuable opportunity to support clinical research by containing crucial patient information, including diagnoses, symptoms, medications, lab tests, and more. Despite the success of deep learning for biomedical Named Entity Recognition (NER), the literature in this field still presents a gap regarding applications focused on lung cancer for the Italian language. Hence, this paper presents a transformer-based approach to extract named entities from Italian clinical notes related to Non-Small Cell Lung Cancer (NSCLC). We introduce a novel set of 25 clinical entities related to NSCLC building a corpus annotated for NER. We apply a state-of the-art model pre-trained on Italian biomedical texts to the manually annotated clinical reports of a cohort of 257 patients suffering from NSCLC, successfully dealing with class-imbalance problems and obtaining promising performance (average F1-score of 84.3%). We also compared our method with two other pre-trained state-of-the-art models showing that the domain specific knowledge offered by the proposed approach is necessary to achieve higher performance. These findings also showcase the feasibility of using transformers to extract biomedical information in the Italian language.

The use of Machine Learning in IoT devices has become the only viable path in today's landscape, where millions of connected devices surround us and increasingly affect our lives. These resource-limited devices interact with the surrounding world via actuators and sensors. Many of these devices use Machine Learning techniques to be able to interpret the world and choose the appropriate action to take. Therefore the purpose of this work is to create a system that allows the application of Machine Learning algorithms directly to the ends of the network, where sensors and actuators reside. The system is designed to rely on the SENSIPLUS smart-sensor as a data acquisition device, and consists of an automatic code generation and compilation system, which through the use of a Toolchain, allows to run artificial intelligence algorithms directly on microcontroller devices.

Water monitoring systems continuously working ensure real–time pollutant detection capabilities according to their sensitivity and specificity. It is necessary to balance such features because, although being able to sense several substances is a desired feature, the reduction of false positives is a primary goal a classification system should have. High false positive makes the system unusable. The current solution enables a 24/7 service with a sampling rate equal to 0.6 Hz. Our goal is to limit false positives to 1 per day, thus achieving 99.99% accuracy at least. In this paper, we add a false positive reduction module to our pre-existent system, aiming to manage false positive boosters as sensor drift and signal oscillations. Obtained results, using a Multi Layer Perceptron classifier, confirm the false positive reduction while keeping high true positive rates.

In smart city framework, the water monitoring through an efficient, low-cost, low-power and IoT-oriented sensor technology is a crucial aspect to allow, with limited resources, the analysis of contaminants eventually affecting wastewater. In this sense, common interfering substances, as detergents, cannot be classified as dangerous contaminants and should be neglected in the classification. By adopting classical machine learning approaches having a finite set of possible responses, each alteration of the sensor baseline is always classified as one out of the predetermined substances. Consequently, we developed an anomaly detection system based on one-class classifiers, able to discriminate between a recognized set of substances and an interfering source. In this way, the proposed detection system is able to provide detailed information about the water status and distinguish between harmless detergents and dangerous contaminants.

Nowadays water monitoring represents one of the most challenging global aims for the protection of people and environment health. In this paper we propose the application of an integrated system for the detection and recognition of contaminants in water. It is based on a two layer architecture: a sensing layer based on SENSIPLUS chip, and a data collection and classification layer, hereafter referred as SENSIPLUS Deep Machine (SDM). The SDM includes: a Micro Controller Unit (MCU), an optional host controller (e.g. laptop, smartphone, etc.) and different software components for data communication, analysis, and classification/regression based on machine learning techniques. Although the SDM classification/regression module can be potentially developed with any machine learning solution, in this paper we adopted an Artificial Neural Network with only one hidden layer to have a lightweight solution suitable to run (for inference) on ultra low power MCU. Aiming at further minimizing the network complexity, two alternative training sessions have been pursued: the first one using raw sensors' data and the second one applying a feature space dimensionality reduction through the Principal Component Analysis technique. Comparable and positive results (higher than 82% as average accuracy) have been obtained, confirming the validity and potentiality of the proposed system.

Image dehazing deals with the removal of undesired loss of visibility in outdoor images due to the presence of fog. Retinex is a color vision model mimicking the ability of the Human Visual System to robustly discount varying illuminations when observing a scene under different spectral lighting conditions. Retinex has been widely explored in the computer vision literature for image enhancement and other related tasks. While these two problems are apparently unrelated, the goal of this work is to show that they can be connected by a simple linear relationship. Specifically, most Retinex-based algorithms have the characteristic feature of always increasing image brightness, which turns them into ideal candidates for effective image dehazing by directly applying Retinex to a hazy image whose intensities have been inverted. In this paper, we give theoretical proof that Retinex on inverted intensities is a solution to the image dehazing problem. Comprehensive qualitative and quantitative results indicate that several classical and modern implementations of Retinex can be transformed into competing image dehazing algorithms performing on pair with more complex fog removal methods, and can overcome some of the main challenges associated with this problem.

In this paper, we analyze how stabilizing the variance of intensity-dependent quantum noise in digital mammograms can significantly improve the computerized detection of microcalcifications (MCs). These lesions appear on mammograms as tiny deposits of calcium smaller than 20 pixels in diameter. At this scale, high frequency image noise is dominated by quantum noise, which in raw mammograms can be described with a square-root noise model. Under this assumption, we derive an adaptive variance stabilizing transform (VST) that stabilizes the noise to unitary standard deviation in all the images. This is achieved by estimating the noise characteristics from the image at hand. We tested the adaptive VST as a preprocessing stage for four existing computerized MC detection methods on three data sets acquired with mammographic units from different manufacturers. In all the test cases considered, MC detection performance on transformed mammograms was statistically significantly higher than on unprocessed mammograms. Results were also superior in comparison with a "fixed" (nonparametric) VST previously proposed for digital mammograms.

Microcalcifications are an early mammographic indicator of breast cancer. To assist screening radiologists in reading mammograms, machine learning techniques have been developed for the automated detection of microcalcifications. In the last few years, Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in many computer vision and medical image analysis applications. A key step in CNN-based detection is image preprocessing, including brightness and contrast variations. In this work, we investigate the influence of preprocessing of digital mammograms on the microcalcification detection performance of two CNNs inspired by the popular AlexNet and VGGnet. We tested two preprocessing methods commonly applied to unprocessed raw digital mammograms: (i) the logarithmic transformation adopted by different manufacturers for the presentation of the image to the radiologists; and (ii) the square-root of image intensity that stabilizes the intensity-dependent noise present in the mammogram. Experiments were performed on 1,066 mammograms acquired with GE Senographe systems. Both preprocessing methods yielded statistically significantly better microcalcification detection performance. Results of the square-root transform were superior to those obtained with the log transform.

Assessment of retinal vessels is fundamental for the diagnosis of many disorders such as heart diseases, diabetes and hypertension. The imaging of retina using advanced fundus camera has become a standard in computer-assisted diagnosis of opthalmic disorders. Modern cameras produce high quality color digital images, but during the acquisition process the light reflected by the retinal surface generates a luminosity and contrast variation. Irregular illumination can introduce severe distortions in the resulting images, decreasing the visibility of anatomical structures and consequently demoting the performance of the automated segmentation of these structures. In this paper, a novel approach for illumination correction of color fundus images is proposed and applied as preprocessing step for retinal vessel segmentation. Our method builds on the connection between two different phenomena, shadows and haze, and works by removing the haze from the image in the inverted intensity domain. This is shown to be equivalent to correct the nonuniform illumination in the original intensity domain. We tested the proposed method as preprocessing stage of two vessel segmentation methods, one unsupervised based on mathematical morphology, and one supervised based on deep learning Convolutional Neural Networks (CNN). Experiments were performed on the publicly available retinal image database DRIVE. Statistically significantly better vessel segmentation performance was achieved in both test cases when illumination correction was applied.

BACKGROUND: Further advances in modern microscopy are leading to teravoxel-sized tiled 3D images at high resolution, thus increasing the dimension of the stitching problem of at least two orders of magnitude. The existing software solutions do not seem adequate to address the additional requirements arising from these datasets, such as the minimization of memory usage and the need to process just a small portion of data. RESULTS: We propose a free and fully automated 3D Stitching tool designed to match the special requirements coming out of teravoxel-sized tiled microscopy images that is able to stitch them in a reasonable time even on workstations with limited resources. The tool was tested on teravoxel-sized whole mouse brain images with micrometer resolution and it was also compared with the state-of-the-art stitching tools on megavoxel-sized publicy available datasets. This comparison confirmed that the solutions we adopted are suited for stitching very large images and also perform well on datasets with different characteristics. Indeed, some of the algorithms embedded in other stitching tools could be easily integrated in our framework if they turned out to be more effective on other classes of images. To this purpose, we designed a software architecture which separates the strategies that use efficiently memory resources from the algorithms which may depend on the characteristics of the acquired images. CONCLUSIONS: TeraStitcher is a free tool that enables the stitching of Teravoxel-sized tiled microscopy images even on workstations with relatively limited resources of memory (<8 GB) and processing power. It exploits the knowledge of approximate tile positions and uses ad-hoc strategies and algorithms designed for such very large datasets. The produced images can be saved into a multiresolution representation to be efficiently retrieved and processed. We provide TeraStitcher both as standalone application and as plugin of the free software Vaa3D.