Human Expert Analysis Research Articles

Detecting APS failures using LSTM-AE and anomaly transformer enhanced with human expert analysis

This study develops a novel semi-supervised approach for detecting Air Pressure System (APS) failures in Heavy-Duty Vehicles (HDVs) by exploiting two modern Machine Learning (ML) models: Long Short-Term Memory Autoencoder (LSTM-AE) and Transformer for Anomaly Detection (TranAD), and enhancing their performance with human expertise. To tackle the failure detection problem, a dataset comprising 30 days of operational time-series data from 110 healthy vehicles with no recorded APS issues and 30 vehicles that experienced APS failures requiring road assistance was acquired. Several preprocessing steps are proposed and three key features are extracted as APS health indicators. These features are then utilized both in human expert analysis (HEA) and training of ML models. When compared to HEA, both LSTM-AE and TranAD models exhibit superior performance individually in APS failure detection, achieving F1 scores of 0.75 and 0.79 respectively, and the same accuracy of 91.4%. Further, the integration of HEA with those ML models enhances model effectiveness in all experimental results, especially in reducing false alarms that cause customer dissatisfaction. The TranAD model combined with human expert analysis achieved the best performance with an unprecedented 0.82 F1 score and 92.8% accuracy. In addition to presenting a new methodology for failure detection, this paper suggests a way for more efficient and reliable predictive maintenance practices for HDVs.

P-101 Machine and Deep learning models to classify Comet assay tests for sperm DNA fragmentation evaluation

Abstract Study question To evaluate the performance of Machine and Deep learning classification models for sperm DNA fragmentation testing. Summary answer Machine and Deep learning models achieved an accuracy over 90% for sperm DNA fragmentation testing. What is known already Artificial Intelligence (AI) is an up-to-date tool that could improve current diagnostics in reproductive medicine. Paternal genome integrity is essential after fertilization. Therefore, high levels of DNA damage have been associated with male infertility. Sperm DNA fragmentation (SDF) is a well-known male fertility biomarker. Among different diagnostic tests analysing SDF, the Comet assay offers high specificity and sensibility. Human expert-analysis has shown to be highly precise but can be influenced by external factors, samples’ variability or tiredness. Consequently, computer-assisted diagnostic based on AI models may be helpful to improve the Comet assay analysis. Study design, size, duration From February to June 2021, alkaline and neutral Comet assays tests were performed on semen samples from a heterogenous group of men. The CometAssay IVTM software was used to record quantitative data from spermatozoa: 500 normal and altered spermatozoa were assessed after the alkaline and neutral Comet assay. Images for each spermatozoon were also taken. The total size was 2000 analysed spermatozoa. Finally, a first validation step was performed using new data from 20000 spermatozoa. Participants/materials, setting, methods For each spermatozoon, ten quantitative parameters and a grayscale image were obtained. A Machine learning predictive model using the Random Forest (RF) algorithm was trained with the quantitative parameters. Moreover, a Deep learning Convolutional Neural Network (CNN) algorithm was trained with cell images. Both models were trained on 67% of data and tested using the remaining 33%. Main results and the role of chance Predictive models based on RF and CNN showed high performance for normal/altered cells automatic classification. The accuracy achieved by the RF models was 95.51% for the alkaline Comet assay and 92.64% for the neutral Comet assay. Regarding the CNN models, the accuracy was 96.71% for the alkaline Comet assay and 93.19% for the neutral Comet assay. CNN models showed better accuracy on both assays. Regarding the quantitative parameters considered in the RF model, the most important parameters for classification are the following ones: Mean_Grey_Level and Total_Intensity in the alkaline Comet assay, and Tail_Migration and Tail_Length in the neutral Comet assay. Finally, 20000 spermatozoa from 100 semen samples were analysed to compare the AI models result with the annotation from an expert human. The Kruskal-Wallis test did not show significant differences for the alkaline and the neutral Comet assays (p > 0.05 for all cases). Paired comparisons using the Mann-Whitney U test did not show statistical differences (p > 0.05 for all cases). According to these results, AI models may reproduce a human analysis. To facilitate the use in the laboratory of the obtained models, a web application was developed to process new samples. Limitations, reasons for caution Diagnostic assays based on continuous variables include a threshold value to separate normal and altered populations. The final result of some samples could be mistaken when a high number of spermatozoa present a fragmentation index near the cut-off value. Wider implications of the findings Diagnostic of male infertility through the analysis of SDF can be achieved through AI predictive models. This technology might help in the standardization of SDF testing between laboratories. Trial registration number not applicable

A Taxonomy of Food Supply Chain Problems from a Computational Intelligence Perspective.

In the last few years, the Internet of Things, and other enabling technologies, have been progressively used for digitizing Food Supply Chains (FSC). These and other digitalization-enabling technologies are generating a massive amount of data with enormous potential to manage supply chains more efficiently and sustainably. Nevertheless, the intricate patterns and complexity embedded in large volumes of data present a challenge for systematic human expert analysis. In such a data-driven context, Computational Intelligence (CI) has achieved significant momentum to analyze, mine, and extract the underlying data information, or solve complex optimization problems, striking a balance between productive efficiency and sustainability of food supply systems. Although some recent studies have sorted the CI literature in this field, they are mainly oriented towards a single family of CI methods (a group of methods that share common characteristics) and review their application in specific FSC stages. As such, there is a gap in identifying and classifying FSC problems from a broader perspective, encompassing the various families of CI methods that can be applied in different stages (from production to retailing) and identifying the problems that arise in these stages from a CI perspective. This paper presents a new and comprehensive taxonomy of FSC problems (associated with agriculture, fish farming, and livestock) from a CI approach; that is, it defines FSC problems (from production to retail) and categorizes them based on how they can be modeled from a CI point of view. Furthermore, we review the CI approaches that are more commonly used in each stage of the FSC and in their corresponding categories of problems. We also introduce a set of guidelines to help FSC researchers and practitioners to decide on suitable families of methods when addressing any particular problems they might encounter. Finally, based on the proposed taxonomy, we identify and discuss challenges and research opportunities that the community should explore to enhance the contributions that CI can bring to the digitization of the FSC.

Informing nuclear physics via machine learning methods with differential and integral experiments

Background: Information from differential nuclear-physics experiments and theory is often too uncertain to accurately define nuclear-physics observables such as cross sections or energy spectra. Integral experimental data, representing the applications of these observables, are often more precise but depend simultaneously on too many of them to unambiguously identify issues in the observable with human expert analysis alone.Purpose: We explore how we can leverage physics knowledge gained from differential experimental data, nuclear theory, integral experiments, and neutron-transport calculations to better understand nuclear-physics observables in the context of the application area represented by integral experiments. We support this task with machine-learning methods to discern trends in a large amount of convoluted data.Methods: Differential and integral information was used in an analysis augmented by the random forest and the Shapley additive explanations metric. We chose as an application area one that is represented by criticality measurements and pulsed-sphere neutron-leakage spectra.Results: We show one representative example ($^{241}\mathrm{Pu}$ fission observables) where the combination of differential and integral information allowed to resolve issues in data representing these observables. As a starting point, the machine learning (ML) algorithms highlighted several observables as leading potentially to bias in simulating integral experiments. Differential information, paired with sensitivity to integral quantities, allowed us then to pinpoint one specific observable ($^{241}\mathrm{Pu}$ fission cross section) as the main driver of bias. The comparison to integral experiments, on the other hand, allowed us to indicate a likely reliable experiment among several discrepant ones for this observables. In other cases (e.g., $^{239}\mathrm{Pu}$ observables), we were not able to resolve the confounding introduced by integral experiments but instead highlighted the need for targeted new experiments and theory developments to better constrain the nuclear-physics space for the application area represented by integral experiments.Conclusions: We were able to combine information from differential experimental data, nuclear-physics theory, integral experiments, and neutron-transport simulations of the latter experiments with the help of the random forest algorithm and expert judgment. This combination of knowledge allows to improve our description of nuclear-physics observables as applied to a particular application area.

On the Feasibility of Adversarial Sample Creation Using the Android System API

Due to its popularity, the Android operating system is a critical target for malware attacks. Multiple security efforts have been made on the design of malware detection systems to identify potentially harmful applications. In this sense, machine learning-based systems, leveraging both static and dynamic analysis, have been increasingly adopted to discriminate between legitimate and malicious samples due to their capability of identifying novel variants of malware samples. At the same time, attackers have been developing several techniques to evade such systems, such as the generation of evasive apps, i.e., carefully-perturbed samples that can be classified as legitimate by the classifiers. Previous work has shown the vulnerability of detection systems to evasion attacks, including those designed for Android malware detection. However, most works neglected to bring the evasive attacks onto the so-called problem space, i.e., by generating concrete Android adversarial samples, which requires preserving the app’s semantics and being realistic for human expert analysis. In this work, we aim to understand the feasibility of generating adversarial samples specifically through the injection of system API calls, which are typical discriminating characteristics for malware detectors. We perform our analysis on a state-of-the-art ransomware detector that employs the occurrence of system API calls as features of its machine learning algorithm. In particular, we discuss the constraints that are necessary to generate real samples, and we use techniques inherited from interpretability to assess the impact of specific API calls to evasion. We assess the vulnerability of such a detector against mimicry and random noise attacks. Finally, we propose a basic implementation to generate concrete and working adversarial samples. The attained results suggest that injecting system API calls could be a viable strategy for attackers to generate concrete adversarial samples. However, we point out the low suitability of mimicry attacks and the necessity to build more sophisticated evasion attacks.

Diagnosis of diabetic retinopathy based on holistic texture and local retinal features

In this paper, eye fundus images are analyzed for the automatic detection of diabetic retinopathy. One thousand two hundred eye fundus images of the Messidor database were used to test the system using the cross validation in various settings. Two types of features were extracted including the holistic texture features and the local retinal features. Four classifiers were implemented including the k-nearest neighbors, neural networks, support vector machines, and random decision forests. The best results from the analysis of holistic texture features were obtained for the Independent Component Analysis method, which had never been tested before in this type of image. Furthermore, the performance of our system improved greatly when two local retinal features — micro-aneurysms and exudates — were incorporated into the analysis, a methodology inspired by a modular approach originally developed for face-recognition tasks. The diagnostic performance of our algorithm is very promising and similar to previous automatic systems and human expert analysis on the same dataset. This framework has the potential to be used as an aiding tool for the diagnosis of diabetic retinopathy.

Semi-Supervised Approach to Phase Identification from Combinatorial Sample Diffraction Patterns

Manual attribution of crystallographic phases from high-throughput x-ray diffraction studies is an arduous task, and represents a rate-limiting step in high-throughput exploration of new materials. Here, we demonstrate a semi-supervised machine learning technique, SS-AutoPhase, which uses a two-step approach to identify automatically phases from diffraction data. First, clustering analysis is used to select a representative subset of samples automatically for human analysis. Second, an AdaBoost classifier uses the labeled samples to identify the presence of the different phases in diffraction data. SS-AutoPhase was used to identify the metallographic phases in 278 diffraction patterns from a FeGaPd composition spread sample. The accuracy of SS-AutoPhase was >82.6% for all phases when 15% of the diffraction patterns were used for training. The SS-AutoPhase predicted phase diagram showed excellent agreement with human expert analysis. Furthermore it was able to determine and identify correctly a previously unreported phase.

Knowledge-based generalization of metabolic models.

Genome-scale metabolic model reconstruction is a complicated process beginning with (semi-)automatic inference of the reactions participating in the organism's metabolism, followed by many iterations of network analysis and improvement. Despite advances in automatic model inference and analysis tools, reconstruction may still miss some reactions or add erroneous ones. Consequently, a human expert's analysis of the model will continue to play an important role in all the iterations of the reconstruction process. This analysis is hampered by the size of the genome-scale models (typically thousands of reactions), which makes it hard for a human to understand them. To aid human experts in curating and analyzing metabolic models, we have developed a method for knowledge-based generalization that provides a higher-level view of a metabolic model, masking its inessential details while presenting its essential structure. The method groups biochemical species in the model into semantically equivalent classes based on the ChEBI ontology, identifies reactions that become equivalent with respect to the generalized species, and factors those reactions into generalized reactions. Generalization allows curators to quickly identify divergences from the expected structure of the model, such as alternative paths or missing reactions, that are the priority targets for further curation. We have applied our method to genome-scale yeast metabolic models and shown that it improves understanding by helping to identify both specificities and potential errors.

Visual Search of Neuropil-Enriched RNAs from Brain In Situ Hybridization Data through the Image Analysis Pipeline Hippo-ATESC

MotivationRNA molecules specifically enriched in the neuropil of neuronal cells and in particular in dendritic spines are of great interest for neurobiology in virtue of their involvement in synaptic structure and plasticity. The systematic recognition of such molecules is therefore a very important task. High resolution images of RNA in situ hybridization experiments contained in the Allen Brain Atlas (ABA) represent a very rich resource to identify them and have been so far exploited for this task through human-expert analysis. However, software tools that may automatically address the same objective are not very well developed.ResultsIn this study we describe an automatic method for exploring in situ hybridization data and discover neuropil-enriched RNAs in the mouse hippocampus. We called it Hippo-ATESC (Automatic Texture Extraction from the Hippocampal region using Soft Computing). Bioinformatic validation showed that the Hippo-ATESC is very efficient in the recognition of RNAs which are manually identified by expert curators as neuropil-enriched on the same image series. Moreover, we show that our method can also highlight genes revealed by microdissection-based methods but missed by human visual inspection. We experimentally validated our approach by identifying a non-coding transcript enriched in mouse synaptosomes. The code is freely available on the web at http://ibislab.ce.unipr.it/software/hippo/.

Semi-supervised clustering for MR brain image segmentation

Magnetic resonance (MR) brain image segmentation of different anatomical structures or tissue types has become a critical requirement in the diagnosis of neurological diseases. Depending on the availability of the training samples, image segmentation can be either supervised or unsupervised. While supervised learning requires a sufficient amount of labelled training data, which is expensive and time-consuming, unsupervised learning techniques suffer from the problem of local traps. Semi-supervised algorithms that includes prior knowledge into the unsupervised learning can enhance the segmentation process without the need of labelled training data. This paper proposes a method to improve the quality of MR brain tissue segmentation and to accelerate the convergence process. The proposed method is a clustering based semi-supervised classifier that does not need a set of labelled training data and uses less human expert analysis than a supervised approach. The proposed classifier labels the voxels clusters of an image slice and then uses statistics and class labels information of the resultant clusters to classify the remaining image slices by applying Gaussian Mixture Model (GMM). The experimental results show that the proposed semi-supervised approach accelerates the convergence and improves the results accuracy when comparing with the classical GMM approach.

Nonlinear optical imaging of integral membrane protein crystals in lipidic mesophases.

Second-order nonlinear optical imaging of chiral crystals (SONICC) is explored for selective detection of integral membrane protein crystals grown in opaque and turbid environments. High turbidity is a hallmark of membrane protein crystallization due to the extensive use of detergent and/or lipids that often form various mesophases. Detection of crystals in such media by conventional optical methods (e.g., intrinsic UV fluorescence, birefringence, bright-field image analysis, etc.) is often complicated by optical scattering and by the small sizes of the crystals that routinely form. SONICC is shown to be well-suited for this application, by nature of its compatibility with imaging in scattering media and its high selectivity for protein crystals. Bright second harmonic generation (SHG) (up to 18 million counts/s) was observed from even relatively small crystals (5 mum) with a minimal background due to the surrounding lipid mesophase ( approximately 1 thousand counts/s). The low background nature of the resulting protein crystal images permitted the use of a relatively simple, particle counting analysis for preliminary scoring. Comparisons between a particle counting analysis of SONICC images and protocols based on the human expert analysis of conventional bright-field and birefringence images were performed.

Multiple Property Tolerance Analysis for the Evaluation of Missense Mutations

Computational prediction of the impact of a mutation on protein function is still not accurate enough for clinical diagnostics without additional human expert analysis. Sequence alignment-based methods have been extensively used but their results highly depend on the quality of the input alignments and the choice of sequences. Incorporating the structural information with alignments improves prediction accuracy. Here, we present a conservation of amino acid properties method for mutation prediction, Multiple Properties Tolerance Analysis (MuTA), and a new strategy, MuTA/S, to incorporate the solvent accessible surface (SAS) property into MuTA. Instead of combining multiple features by machine learning or mathematical methods, an intuitive strategy is used to divide the residues of a protein into different groups, and in each group the properties used is adjusted. The results for LacI, lysozyme, and HIV protease show that MuTA performs as well as the widely used SIFT algorithm while MuTA/S outperforms SIFT and MuTA by 2%–25% in terms of prediction accuracy. By incorporating the SAS term alone, the alignment dependency of overall prediction accuracy is significantly reduced. MuTA/S also defines a new way to incorporate any structural features and knowledge and may lead to more accurate predictions.