NERONE: The Fast Way to Efficiently Execute Your Deep Learning Algorithm at the Edge.

Semantic segmentation and classification are pivotal in many clinical applications, such as radiation dose quantification and surgery planning. While manually labeling images is highly time-consuming, the advent of Deep Learning (DL) has introduced a valuable alternative. Nowadays, DL models inference is run on Graphics Processing Units (GPUs), which are power-hungry devices, and, therefore, are not the most suited solution in constrained environments where Field Programmable Gate Arrays (FPGAs) become an appealing alternative given their remarkable performance per watt ratio. Unfortunately, FPGAs are hard to use for non-experts, and the creation of tools to open their employment to the computer vision community is still limited. For these reasons, we propose NERONE, which allows end users to seamlessly benefit from FPGA acceleration and energy efficiency without modifying their DL development flows. To prove the capability of NERONE to cover different network architectures, we have developed four models, one for each of the chosen datasets (three for segmentation and one for classification), and we deployed them, thanks to NERONE, on three different embedded FPGA-powered boards achieving top average energy efficiency improvements of 3.4× and 1.9× against a mobile and a datacenter GPU devices, respectively.

Rethinking Theta/Beta Ratio in ADHD through Functional Data Analysis.

ADHD is a neurodevelopmental disorder largely diffused among children and adolescents. The current method of diagnosis is based on agreed clinical literature such as DSM-5, by identifying and evaluating signs of hyperactivity and inattention. Multiple reviews have assessed that EEG is not sufficiently reliable for the diagnosis of ADHD. Theta-Beta Ratio is now the sole EEG parameter considered for analysis, although it is not robust enough to be utilized as a confirmatory technique for diagnosis. In this setting, new objective approaches for reliably classifying neurotypical and ADHD subjects are required. As a result, we suggest a new methodology based on Functional Data Analysis, a statistical class of methods for dealing with curves and functions. The initial stage in our method is to separate frequency bands from the EEG signal using a wavelet decomposition. We next compute the Power Spectral Densities of each of these bands and represent them as mathematical functions via spline interpolation. Finally, the relevance of the collected features is assessed using the Permutation ANOVA test. Using this method, we can detect different patterns in the PSDs of the groups and identify statistically significant features, confirming prior findings in the literature. We validate the features using classification techniques such as Bagging trees, Random Forest, and AdaBoost. The latter reaches the highest accuracy score of 76.65%, confirming the relevance of the extracted features.

Towards an Informed CNN for Bone SR-microCT Image Classification with an Unsupervised Patched-based Image Clustering.

Bone microscale differences cannot be readily recognizable to humans from Synchrotron Radiation micro-Computed Tomography (SR-microCT) images. Premises are possible with Deep Learning (DL) imaging analysis. Despite this, more attention to high-level features leads models to require help identifying relevant details to support a decision. Within this context, we propose a method for classifying healthy, osteoporotic, and COVID-19 femoral heads SR-microCT images informing a vgg16 about the most subtle microscale differences using unsupervised patched-based clustering. Our strategy allows achieving up to 9.8% accuracy improvement in classifying healthy from osteoporotic images over uninformed methods, while 59.1% of accuracy between osteoporosis and COVID-19.Clinical relevance-We established a starting point for classifying healthy, osteoporotic, and COVID-19 femoral heads from SR-microCTs with human non-discriminative features, with 60.91% accuracy in healthy-osteporotic image classification.

Analysis and Classification of Event-Related Potentials During Image Observation.

In the field of cognitive neuroscience, researchers have conducted extensive studies on object categorization using Event-Related Potential (ERP) analysis, specifically by analyzing electroencephalographic (EEG) response signals triggered by visual stimuli. The most common approach for visual ERP analysis is to use a low presentation rate of images and an active task where participants actively discriminate between target and non-target images. However, researchers are also interested in understanding how the human brain processes visual information in real-world scenarios. To simulate real-life object recognition, this study proposes an analysis pipeline of visual ERPs evoked by images presented in a Rapid Serial Visual Presentation (RSVP) paradigm. Such an approach allows for the investigation of recurrent patterns of visual ERP signals across specific categories and subjects. The pipeline includes segmentation of the EEGs in epochs, and the use of the resulting features as inputs for Support Vector Machine (SVM) classification. Results demonstrate common ERP patterns across the selected categories and the ability to obtain discriminant information from single visual stimuli presented in the RSVP paradigm.

Anatomically compliant modes of variations: New tools for brain connectivity.

Anatomical complexity and data dimensionality present major issues when analysing brain connectivity data. The functional and anatomical aspects of the connections taking place in the brain are in fact equally relevant and strongly intertwined. However, due to theoretical challenges and computational issues, their relationship is often overlooked in neuroscience and clinical research. In this work, we propose to tackle this problem through Smooth Functional Principal Component Analysis, which enables to perform dimensional reduction and exploration of the variability in functional connectivity maps, complying with the formidably complicated anatomy of the grey matter volume. In particular, we analyse a population that includes controls and subjects affected by schizophrenia, starting from fMRI data acquired at rest and during a task-switching paradigm. For both sessions, we first identify the common modes of variation in the entire population. We hence explore whether the subjects' expressions along these common modes of variation differ between controls and pathological subjects. In each session, we find principal components that are significantly differently expressed in the healthy vs pathological subjects (with p-values < 0.001), highlighting clearly interpretable differences in the connectivity in the two subpopulations. For instance, the second and third principal components for the rest session capture the imbalance between the Default Mode and Executive Networks characterizing schizophrenia patients.

An Optimized U-Net for Unbalanced Multi-Organ Segmentation.

Medical practice is shifting towards the automation and standardization of the most repetitive procedures to speed up the time-to-diagnosis. Semantic segmentation repre-sents a critical stage in identifying a broad spectrum of regions of interest within medical images. Indeed, it identifies relevant objects by attributing to each image pixels a value representing pre-determined classes. Despite the relative ease of visually locating organs in the human body, automated multi-organ segmentation is hindered by the variety of shapes and dimensions of organs and computational resources. Within this context, we propose BIONET, a U-Net-based Fully Convolutional Net-work for efficiently semantically segmenting abdominal organs. BIONET deals with unbalanced data distribution related to the physiological conformation of the considered organs, reaching good accuracy for variable organs dimension with low variance, and a Weighted Global Dice Score score of 93.74 ± 1.1%, and an inference performance of 138 frames per second. Clinical Relevance - This work established a starting point for developing an automatic tool for semantic segmentation of variable-sized organs within the abdomen, reaching considerable accuracy on small and large organs with low variability, reaching a 93.74 ± 1.1 % of Weighted Global Dice Score.

Math Skills: a New Look from Functional Data Analysis.

Mental calculations involve various areas of the brain. The frontal, parietal and temporal lobes of the left hemisphere have a principal role in the completion of this typology of tasks. Their level of activation varies based on the mathematical competence and attentiveness of the subject under examination and the perceived difficulty of the task. Recent literature often investigates patterns of cerebral activity through fMRI, which is an expensive technique. In this scenario, EEGs represent a more straightforward and cheaper way to collect information regarding brain activity. In this work, we propose an EEG based method to detect differences in the cerebral activation level of people characterized by different abilities in carrying out the same arithmetical task. Our approach consists in the extraction of the activation level of a given region starting from the EEG acquired during resting state and during the completion of a subtraction task. We then analyze these data through Functional Data Analysis, a statistical technique that allows operating on biomedical signals as if they were functions. The application of this technique allowed for the detection of distinct cerebral patterns among the two groups and, more specifically, highlighted the presence of higher levels of activation in the parietal lobe in the population characterized by a lower performance.

On the Automation of Radiomics-Based Identification and Characterization of NSCLC.

Proper detection and accurate characterization of Non-Small Cell Lung Cancer (NSCLC) are an open challenge in the imaging field. Biomedical imaging is fundamental in lung cancer assessment and offers the possibility of calculating predictive biomarkers impacting patients' management. Within this context, radiomics, which consists of extracting quantitative features from digital images, shows encouraging results for clinical applications, but the sub-optimal standardization of the procedure and the lack of definitive results are still a concern in the field. For these reasons, this work proposes the design and development of LuCIFEx, a fully-automated pipeline for non-invasive in-vivo characterization of NSCLC, aiming to speed up the analysis process and enable an early diagnosis of the tumor.LuCIFEx pipeline relies on routinely acquired [18F]FDG-PET/CT images for the automatic segmentation of the cancer lesion, allowing the computation of accurate radiomic features, then employed for cancer characterization through Machine Learning algorithms. The proposed multi-stage segmentation process can identify the lesion with a mean accuracy of 94.2±5.0%. Finally, the proposed data analysis pipeline demonstrates the potential of PET/CT features for the automatic recognition of lung metastases and NSCLC histological subtypes, while highlighting the main current limitations of the radiomic approach.

Effects of a cool classroom microclimate on cardiac autonomic control and cognitive performances in undergraduate students.

An inverted U-shape relationship between cognitive performance and indoor temperature with best performance peaking at 21.6 °C was previously described. Little is known on classroom temperature reduction effects on cognitive performances and cardiac autonomic profile, during the cold season. Fifteen students underwent electrocardiogram recording during a lecture in two days in December when classroom temperatures were set as neutral (NEUTRAL, 20-22 °C) and cool (COOL, 16-18 °C). Cognitive performance (memory, verbal ability, reasoning, overall cognitive C-score) was assessed by Cambridge Brain Science cognitive evaluation tool. Cardiac autonomic control was evaluated via the analysis of spontaneous fluctuations of heart period, as the temporal distance between two successive R-wave peaks (RR). Spectral analysis provided the power in the high frequency (HF, 0.15-0.40 Hz) and low frequency (LF, 0.04-0.15 Hz) bands of RR variability. Sympatho-vagal interaction was assessed by LF to HF ratio (LF/HF). Symbolic analysis provided the fraction of RR patterns composed by three heart periods with no variation (0 V%) and two variations (2 V%), taken as markers of cardiac sympathetic and vagal modulations, respectively. The students' thermal comfort was assessed during NEUTRAL and COOL trials. Classroom temperatures were 21.5 ± 0.8 °C and 18.4 ± 0.4 °C during NEUTRAL and COOL. Memory, verbal ability, C-Score were greater during COOL (13.01 ± 3.43, 12.32 ± 2.58, 14.29 ± 2.90) compared to NEUTRAL (9.98 ± 2.26, p = 0.002; 8.57 ± 1.07, p = 0.001 and 10.35 ± 3.20, p = 0.001). LF/HF (2.4 ± 1.7) and 0 V% (23.2 ± 11.1%) were lower during COOL compared to NEUTRAL (3.7 ± 2.8, p = 0.042; 28.1 ± 12.2.1%, p = 0.031). During COOL, 2 V% was greater (30.5 ± 10.9%) compared to NEUTRAL (26.2 ± 11.3, p = 0.047). The students' thermal comfort was slightly reduced during COOL compared to NEUTRAL trial. During cold season, a better cognitive performance was obtained in a cooler indoor setting enabling therefore energy saving too.

A Functional Data Analysis Approach to Left Ventricular Remodeling Assessment.

Left ventricular remodeling is a mechanism common to various cardiovascular diseases affecting myocardial morphology. It can be often overlooked in clinical practice since the parameters routinely employed in the diagnostic process (e.g., the ejection fraction) mainly focus on evaluating volumetric aspects. Nevertheless, the integration of a quantitative assessment of structural modifications can be pivotal in the early individuation of this pathology. In this work, we propose an approach based on functional data analysis to evaluate myocardial contractility. A functional representation of ventricular shape is introduced, and functional principal component analysis and depth measures are used to discriminate healthy subjects from those affected by left ventricular hypertrophy. Our approach enables the integration of higher informative content compared to the traditional clinical parameters, allowing for a synthetic representation of morphological changes in the myocardium, which could be further explored and considered for future clinical practice implementation.

Development of a Deep-Learning Pipeline to Recognize and Characterize Macrophages in Colo-Rectal Liver Metastasis.

Quantitative analysis of Tumor Microenvironment (TME) provides prognostic and predictive information in several human cancers but, with few exceptions, it is not performed in daily clinical practice since it is extremely time-consuming. We recently showed that the morphology of Tumor Associated Macrophages (TAMs) correlates with outcome in patients with Colo-Rectal Liver Metastases (CLM). However, as for other TME components, recognizing and characterizing hundreds of TAMs in a single histopathological slide is unfeasible. To fasten this process, we explored a deep-learning based solution. We tested three Convolutional Neural Networks (CNNs), namely UNet, SegNet and DeepLab-v3, with three different segmentation strategies, semantic segmentation, pixel penalties and instance segmentation. The different experiments are compared according to the Intersection over Union (IoU), a metric describing the similarity between what CNN predicts as TAM and the ground truth, and the Symmetric Best Dice (SBD), which indicates the ability of CNN to separate different TAMs. UNet and SegNet showed intrinsic limitations in discriminating single TAMs (highest SBD 61.34±2.21), whereas DeepLab-v3 accurately recognized TAMs from the background (IoU 89.13±3.85) and separated different TAMs (SBD 79.00±3.72). This deep-learning pipeline to recognize TAMs in digital slides will allow the characterization of TAM-related metrics in the daily clinical practice, allowing the implementation of prognostic tools.

A faster approach to ECG analysis in emergency situations.

Every day, a substantial number of people need to be treated in emergencies and these situations imply a short timeline. Especially concerning heart abnormalities, the time factor is very important. Therefore, we propose a full-stack system for faster and cheaper ECG taking aimed at paramedics, to enhance Emergency Medical Service (EMS) response time. To stick with the golden hour rule, and reduce the cost of the current devices, the system is capable of enabling the detection and annotation of anomalies during ECG acquisition. Our system combines Machine Learning and traditional Signal Processing techniques to analyze ECG tracks to use it in a glove-like wearable. Finally, a graphical interface offers a dynamic view of the whole procedure.

GPU-accelerated Chemical Similarity Assessment for Large Scale Databases.

The assessment of chemical similarity between molecules is a basic operation in chemoinformatics, a computational area concerning with the manipulation of chemical structural information. Comparing molecules is the basis for a wide range of applications such as searching in chemical databases, training prediction models for virtual screening or aggregating clusters of similar compounds. However, currently available multimillion databases represent a challenge for conventional chemoinformatics algorithms raising the necessity for faster similarity methods. In this paper, we extensively analyze the advantages of using many-core architectures for calculating some commonly-used chemical similarity coefficients such as Tanimoto, Dice or Cosine. Our aim is to provide a wide-breath proof-of-concept regarding the usefulness of GPU architectures to chemoinformatics, a class of computing problems still uncovered. In our work, we present a general GPU algorithm for all-to-all chemical comparisons considering both binary fingerprints and floating point descriptors as molecule representation. Subsequently, we adopt optimization techniques to minimize global memory accesses and to further improve efficiency. We test the proposed algorithm on different experimental setups, a laptop with a low-end GPU and a desktop with a more performant GPU. In the former case, we obtain a 4-to-6-fold speed-up over a single-core implementation for fingerprints and a 4-to-7-fold speed-up for descriptors. In the latter case, we respectively obtain a 195-to-206-fold speed-up and a 100-to-328-fold speed-up.