Composite Dataset Research Articles

MSeg: A Composite Dataset for Multi-Domain Semantic Segmentation.

We present MSeg, a composite dataset that unifies semantic segmentation datasets from different domains. A naive merge of the constituent datasets yields poor performance due to inconsistent taxonomies and annotation practices. We reconcile the taxonomies and bring the pixel-level annotations into alignment by relabeling more than 220,000 object masks in more than 80,000 images, requiring more than 1.34 years of collective annotator effort. The resulting composite dataset enables training a single semantic segmentation model that functions effectively across domains and generalizes to datasets that were not seen during training. We adopt zero-shot cross-dataset transfer as a benchmark to systematically evaluate a model's robustness and show that MSeg training yields substantially more robust models in comparison to training on individual datasets or naive mixing of datasets without the presented contributions. A model trained on MSeg ranks first on the WildDash-v1 leaderboard for robust semantic segmentation, with no exposure to WildDash data during training. We evaluate our models in the 2020 Robust Vision Challenge (RVC) as an extreme generalization experiment. MSeg training sets include only three of the seven datasets in the RVC; more importantly, the evaluation taxonomy of RVC is different and more detailed. Surprisingly, our model shows competitive performance and ranks second. To evaluate how close we are to the grand aim of robust, efficient, and complete scene understanding, we go beyond semantic segmentation by training instance segmentation and panoptic segmentation models using our dataset. Moreover, we also evaluate various engineering design decisions and metrics, including resolution and computational efficiency. Although our models are far from this grand aim, our comprehensive evaluation is crucial for progress. We share all the models and code with the community.

Machine learning models to predict micronutrient profile in food after processing

The information on nutritional profile of cooked foods is important to both food manufacturers and consumers, and a major challenge to obtaining precise information is the inherent variation in composition across biological samples of any given raw ingredient. The ideal solution would address precision and generability, but the current solutions are limited in their capabilities; analytical methods are too costly to scale, retention-factor based methods are scalable but approximate, and kinetic models are bespoke to a food and nutrient. We provide an alternate solution that predicts the micronutrient profile in cooked food from the raw food composition, and for multiple foods. The prediction model is trained on an existing food composition dataset and has a 31% lower error on average (across all foods, processes and nutrients) than predictions obtained using the baseline method of retention-factors. Our results argue that data scaling and transformation prior to training the models is important to mitigate any yield bias. This study demonstrates the potential of machine learning methods over current solutions, and additionally provides guidance for the future generation of food composition data, specifically for sampling approach, data quality checks, and data representation standards.

Intelligent Rolling Bearing Fault Diagnosis Method Using Symmetrized Dot Pattern Images and CBAM-DRN

Rolling bearings are a vital component of mechanical equipment. It is crucial to implement rolling bearing fault diagnosis research to guarantee the stability of the long-term action of mechanical equipment. Conversion of rolling bearing vibration signals into images for fault diagnosis research has been a practical diagnostic approach. The current paper presents a rolling bearing fault diagnosis method using symmetrized dot pattern (SDP) images and a deep residual network with convolutional block attention module (CBAM-DRN). The rolling bearing vibration signal is first visualized and transformed into an SDP image with distinct fault characteristics. Then, CBAM-DRN is utilized to derive characteristics directly and detect faults from the input SDP images. In order to prevent conventional time-frequency images from being limited by their inherent flaws and avoid missing the fault features, the SDP technique is employed to convert vibration signals into images for visualization. DRN enables adequate extraction of rolling bearing fault characteristics and prevents training difficulties and gradient vanishing in deep level networks. CBAM assists the diagnostic model in concentrating on the image's more distinctive parts and preventing the interference of non-featured parts. Finally, the method's validity was tested with a composite fault dataset of motor bearings containing multiple loads and fault diameters. The experimental results reflect that the presented approach can attain a diagnostic precision of over 99% and good stability and generalization.

Comparison of template-based and multilayer perceptron-based approach for automatic question generation system

Intelligent tutoring systems are computer-assisted learning systems with adaption to students using artificial intelligence tools. An intelligent tutoring system can drastically improve education efficiency as it provides solutions to many issues that now plague the educational industry. One important component in education is questioning learners to assess and reinforce learning. This research compares two approaches for automatic question generation, a template-based question generation strategy and the phrase-Level automatic question generation system utilizing the Multilayer perceptron model. A template-based technique is a baseline for automatic question generation that uses templates taken from the training set to generate questions by filling certain templates with specific topic items. We utilize question-answer sentence composition datasets and manually constructed datasets for our experiments and comparison of the Multilayer perceptron training model. We used both human and automatic evaluation metrics to assess the efficiency of our suggested methods. Regarding automatic metrics, we selected the BLEU-n gram and ROUGE-N methodologies. The evaluation results demonstrate that the phrase-level Multilayer perceptron-based strategy dominates the template-based approach and has a promising score in both ROUGE automatic and human evaluation metrics.

Monitoring Short-Term Morphobathymetric Change of Nearshore Seafloor Using Drone-Based Multispectral Imagery

Short-term changes in shallow bathymetry affect the coastal zone, and therefore their monitoring is an essential task in coastal planning projects. This study provides a novel approach for monitoring shallow bathymetry changes based on drone multispectral imagery. Particularly, we apply a shallow water inversion algorithm on two composite multispectral datasets, being acquired five months apart in a small Mediterranean sandy embayment (Chania, Greece). Initially, we perform radiometric corrections using proprietary software, and following that we combine the bands from standard and multispectral cameras, resulting in a six-band composite image suitable for applying the shallow water inversion algorithm. Bathymetry inversion results showed good correlation and low errors (<0.3 m) with sonar measurements collected with an uncrewed surface vehicle (USV). Bathymetry maps and true-color orthomosaics assist in identifying morphobathymetric features representing crescentic bars with rip channel systems. The temporal bathymetry and true-color data reveal important erosional and depositional patterns, which were developed under the impact of winter storms. Furthermore, bathymetric profiles show that the crescentic bar appears to migrate across and along-shore over the 5-months period. Drone-based multispectral imagery proves to be an important and cost-effective tool for shallow seafloor mapping and monitoring when it is combined with shallow water analytical models.

Host biology, ecology and the environment influence microbial biomass and diversity in 101 marine fish species

Fish are the most diverse and widely distributed vertebrates, yet little is known about the microbial ecology of fishes nor the biological and environmental factors that influence fish microbiota. To identify factors that explain microbial diversity patterns in a geographical subset of marine fish, we analyzed the microbiota (gill tissue, skin mucus, midgut digesta and hindgut digesta) from 101 species of Southern California marine fishes, spanning 22 orders, 55 families and 83 genera, representing ~25% of local marine fish diversity. We compare alpha, beta and gamma diversity while establishing a method to estimate microbial biomass associated with these host surfaces. We show that body site is the strongest driver of microbial diversity while microbial biomass and diversity is lowest in the gill of larger, pelagic fishes. Patterns of phylosymbiosis are observed across the gill, skin and hindgut. In a quantitative synthesis of vertebrate hindguts (569 species), we also show that mammals have the highest gamma diversity when controlling for host species number while fishes have the highest percent of unique microbial taxa. The composite dataset will be useful to vertebrate microbiota researchers and fish biologists interested in microbial ecology, with applications in aquaculture and fisheries management.

Are plankton nets a thing of the past? An assessment of in situ imaging of zooplankton for large-scale ecosystem assessment and policy decision-making

Zooplankton are fundamental to aquatic ecosystem services such as carbon and nutrient cycling. Therefore, a robust evidence base of how zooplankton respond to changes in anthropogenic pressures, such as climate change and nutrient loading, is key to implementing effective policy-making and management measures. Currently, the data on which to base this evidence, such as long time-series and large-scale datasets of zooplankton distribution and community composition, are too sparse owing to practical limitations in traditional collection and analysis methods. The advance of in situ imaging technologies that can be deployed at large scales on autonomous platforms, coupled with artificial intelligence and machine learning (AI/ML) for image analysis, promises a solution. However, whether imaging could reasonably replace physical samples, and whether AI/ML can achieve a taxonomic resolution that scientists trust, is currently unclear. We here develop a roadmap for imaging and AI/ML for future zooplankton monitoring and research based on community consensus. To do so, we determined current perceptions of the zooplankton community with a focus on their experience and trust in the new technologies. Our survey revealed a clear consensus that traditional net sampling and taxonomy must be retained, yet imaging will play an important part in the future of zooplankton monitoring and research. A period of overlapping use of imaging and physical sampling systems is needed before imaging can reasonably replace physical sampling for widespread time-series zooplankton monitoring. In addition, comprehensive improvements in AI/ML and close collaboration between zooplankton researchers and AI developers are needed for AI-based taxonomy to be trusted and fully adopted. Encouragingly, the adoption of cutting-edge technologies for zooplankton research may provide a solution to maintaining the critical taxonomic and ecological knowledge needed for future zooplankton monitoring and robust evidence-based policy decision-making.

A complex patchwork of magma bodies that fed the Bishop Tuff supereruption (Long Valley Caldera, CA, United States): Evidence from matrix glass major and trace-element compositions

The geologic record reveals events in which enormous volumes (100–1000s of km3) of magma were erupted in a matter of days to months. Yet, the architecture of magmatic systems that feed supereruptions can only be investigated through the study of ancient systems. For more than 40 years, the Bishop Tuff (Long Valley, California) has been the archetypal example of a single, zoned magma body that fed a supereruption. Early-erupted material is pyroxene-free and crystal poor (< 20 wt%), presumably erupted from the upper parts of the magma body; late-erupted material is orthopyroxene and clinopyroxene-bearing, commonly more crystal rich (up to 30 wt% crystals), and presumably tapped magma from the lower portions of the magma body. Fe-Ti oxide compositions suggest higher crystallization temperatures for late-erupted magmas (as high as 820°C) than for early-erupted magmas (as low as 700°C). Pressures derived from major-element compositions of glass inclusions were used to suggest an alternative model of lateral juxtaposition of two main magma bodies—each one feeding early-erupted and late-erupted units. Yet, this interpretation has proven controversial. We present a large dataset of matrix glass compositions for 227 pumice clasts that span the stratigraphy of the deposit. We calculate crystallization pressures based on major-element glass compositions using rhyolite-MELTS geobarometry and crystallization temperatures based on Zr in glass using zircon-saturation geothermometry. Additionally, we apply the same methods to 1,538 major-element and 615 trace-element analyses from a dataset from the literature. The results overwhelmingly demonstrate that the variations in crystallization temperature and pressure are not consistent with vertical stratification of a single magma body. All crystallization pressures and temperatures are very similar, with modes of ∼150 MPa and ∼730°C. Our results support lateral juxtaposition of three main magma bodies. Magmas represented by smaller stratigraphic units crystallized at similar pressures as the main bodies, which suggests coexistence of larger and smaller magma bodies at the time of eruption. We compare our findings with results for other very large eruptions and supereruptions. We argue that supereruptions typically mobilize a complex patchwork of magma bodies that reside within specific levels of the crust. These eruptions reveal the architecture of the crust during moments of high abundance of eruptible magma, revealing crustal states that differ from what is inferred for magmatic systems currently present on Earth.

Micro-expression recognition model based on TV-L1 optical flow method and improved ShuffleNet

Micro-expression is a kind of facial action that reflects the real emotional state of a person, and has high objectivity in emotion detection. Therefore, micro-expression recognition has become one of the research hotspots in the field of computer vision in recent years. Research with neural networks with convolutional structure is still one of the main methods of recognition. This method has the advantage of high operational efficiency and low computational complexity, but the disadvantage is its localization of feature extraction. In recent years, there are more and more plug-and-play self-attentive modules being used in convolutional neural networks to improve the ability of the model to extract global features of the samples. In this paper, we propose the ShuffleNet model combined with a miniature self-attentive module, which has only 1.53 million training parameters. First, the start frame and vertex frame of each sample will be taken out, and its TV-L1 optical flow features will be extracted. After that, the optical flow features are fed into the model for pre-training. Finally, the weights obtained from the pre-training are used as initialization weights for the model to train the complete micro-expression samples and classify them by the SVM classifier. To evaluate the effectiveness of the method, it was trained and tested on a composite dataset consisting of CASMEII, SMIC, and SAMM, and the model achieved competitive results compared to state-of-the-art methods through cross-validation of leave-one-out subjects.

Focusing on your subject: Deep subject-aware image composition recommendation networks

Photo composition is one of the most important factors in the aesthetics of photographs. As a popular application, composition recommendation for a photo focusing on a specific subject has been ignored by recent deep-learning-based composition recommendation approaches. In this paper, we propose a subject-aware image composition recommendation method, SAC-Net, which takes an RGB image and a binary subject window mask as input, and returns good compositions as crops containing the subject. Our model first determines candidate scores for all possible coarse cropping windows. The crops with high candidate scores are selected and further refined by regressing their corner points to generate the output recommended cropping windows. The final scores of the refined crops are predicted by a final score regression module. Unlike existing methods that need to preset several cropping windows, our network is able to automatically regress cropping windows with arbitrary aspect ratios and sizes. We propose novel stability losses for maximizing smoothness when changing cropping windows along with view changes. Experimental results show that our method outperforms state-of-the-art methods not only on the subject-aware image composition recommendation task, but also for general purpose composition recommendation. We also have designed a multistage labeling scheme so that a large amount of ranked pairs can be produced economically. We use this scheme to propose the first subject-aware composition dataset SACD, which contains 2777 images, and more than 5 million composition ranked pairs. The SACD dataset is publicly available at https://cg.cs.tsinghua.edu.cn/SACD/.

A Novel Attention-Based Lightweight Network for Multiscale Object Detection in Underwater Images

Underwater images have low quality, and underwater targets have different sizes. The mainstream target detection networks cannot achieve good results in detecting objects from underwater images. In this study, a lightweight underwater multiscale target detection model with an attention mechanism is designed to solve the above problems. In this model, MobileNetv3 is used as the backbone network for preliminary feature extraction. The lightweight feature extraction module (LFEM) pays attention to the feature map at the channel and space levels. The features with large weights are promoted, while the features with small weights are suppressed. Meanwhile, cross-group information exchange enriches the semantic information and location information of the objects. The context aggregation module (CIAM) pools the extracted feature maps to obtain feature pyramids, and it uses the upsampling-feature refinement-cascade addition (URC) method to effectively fuse global context information and enhance the feature representation. The scale normalization for feature pyramids (SNFP) performs adaptive multiscale perception and multianchor detection on feature maps to cover objects of different sizes and realize multiscale object detection in underwater images. The proposed network can realize lightweight feature extraction, effectively handle the global relationship between the underwater scene and the object while expanding the receptive field, traverse the objects of different scales, and achieve adaptive multianchor detection of multiscale objects in underwater images. The experimental results indicate that our method achieves an average accuracy of 81.94% and a detection speed of 44.3 FPS on a composite dataset. Also, our method is better than the mainstream object detection networks in terms of detection accuracy, lightweight design, and real-time performance.

Using linkage to assess coverage of population estimates.

ObjectivesThe Demographic Index (DI) comprises of five linked administrative datasets, used for population estimation. Current linkage methods are not ideal to utilise the power of this asset. Using the 2021 England and Wales Census, we are developing an innovative composite linkage method to fully utilise the power of the DI.
 ApproachUsing non-greedy deterministic and probabilistic linkage methods, we will link the DI to the Census at a composite level where we believe links exist – i.e., linking a Census cluster (consisting of linked Census and Census Coverage Survey (CCS) records) with a DI cluster (consisting of linked records from the data sources used to make the DI). We will then conduct a pairwise linkage of records from these linked clusters to link individual source records to the Census. We will utilise clerical review to resolve uncertain and conflicting links and to inform the quality of our linkage.
 ResultsWe anticipate producing a high-quality linkage that will inform how the coverage of the DI compares to Census (through the composite-level linkage) and the quality of the DI itself (through the pairwise-level linkage). We have developed a clerical matching system that can display composite-level linkage, i.e., candidate cluster-pairs. We will tailor our clerical review and quality assessment to records that fall within carefully chosen postcode areas, to ensure all hard-to-count groups and geographical areas are sampled. Working with large datasets is a challenge we are overcoming by using distributed computing and search space reduction.
 The 2021 Census has been previously linked to the CCS with high accuracy; these records are considered intrinsically linked.
 ConclusionTo assess national population estimates’ quality and the policy decisions based upon them, we are linking a key composite population-level dataset to the 2021 England and Wales Census. The presentation will showcase the methods we are developing and how we are ensuring the highest quality possible.

Analysing linked 2021 Census and Admin data to inform population transformation statistics.

ObjectivesTo analyse the linked dataset between England and Wales 2021 Census/Census Coverage Survey and the Demographic Index (DI), would provide insights on quality and coverage, enabling improvements of using admin data to form population/ household estimates for future linkages and to evidence additional admin sources needed to capture the population. ApproachThe DI is a composite dataset made up of several linked admin sources. Therefore, the linked output is complex, made up of clusters where individual sources within the clusters have been linked to the Census using the different available information across sources. To carry out the analysis, a flagging strategy has been designed to enable analysts to form cuts of the linked dataset that are specific to their research needs. High level research questions have been designed to provide fast paced analysis to inform the National Statistician’s 2023 recommendation on the need for a future Census. ResultsThe results will inform the National Statistician’s 2023 recommendation, but also the future delivery of transformed population and migrations statistics . This includes: Providing insight into how to improve the current linkage methodology of the DI How we use the DI to construct Statistical Population Datasets (which use inclusion rules to establish the usual resident population) Identifying the extent of over and under coverage in the Statistical Population Datasets, allowing development of an appropriate estimation strategy to more accurately estimate the population The results will also inform how we utilise linked composite data sources in the future, as we can share the lessons learnt in how we planned to use the data, including our flagging strategy verses the reality faced when undertaking the analysis. ConclusionThis analysis will inform approaches to using linked composite data in the future but also provide a wealth of knowledge including: informing linkage methods, defining the population estimation challenge and providing insights into how we deliver admin based population estimates.

Detecting Fear-Memory-Related Genes from Neuronal scRNA-seq Data by Diverse Distributions and Bhattacharyya Distance.

The detection of differentially expressed genes (DEGs) is one of most important computational challenges in the analysis of single-cell RNA sequencing (scRNA-seq) data. However, due to the high heterogeneity and dropout noise inherent in scRNAseq data, challenges in detecting DEGs exist when using a single distribution of gene expression levels, leaving much room to improve the precision and robustness of current DEG detection methods. Here, we propose the use of a new method, DEGman, which utilizes several possible diverse distributions in combination with Bhattacharyya distance. DEGman can automatically select the best-fitting distributions of gene expression levels, and then detect DEGs by permutation testing of Bhattacharyya distances of the selected distributions from two cell groups. Compared with several popular DEG analysis tools on both large-scale simulation data and real scRNA-seq data, DEGman shows an overall improvement in the balance of sensitivity and precision. We applied DEGman to scRNA-seq data of TRAP; Ai14 mouse neurons to detect fear-memory-related genes that are significantly differentially expressed in neurons with and without fear memory. DEGman detected well-known fear-memory-related genes and many novel candidates. Interestingly, we found 25 DEGs in common in five neuron clusters that are functionally enriched for synaptic vesicles, indicating that the coupled dynamics of synaptic vesicles across in neurons plays a critical role in remote memory formation. The proposed method leverages the advantage of the use of diverse distributions in DEG analysis, exhibiting better performance in analyzing composite scRNA-seq datasets in real applications.

Integrated 3d flow-based multi-atlas brain structure segmentation.

MRI brain structure segmentation plays an important role in neuroimaging studies. Existing methods either spend much CPU time, require considerable annotated data, or fail in segmenting volumes with large deformation. In this paper, we develop a novel multi-atlas-based algorithm for 3D MRI brain structure segmentation. It consists of three modules: registration, atlas selection and label fusion. Both registration and label fusion leverage an integrated flow based on grayscale and SIFT features. We introduce an effective and efficient strategy for atlas selection by employing the accompanying energy generated in the registration step. A 3D sequential belief propagation method and a 3D coarse-to-fine flow matching approach are developed in both registration and label fusion modules. The proposed method is evaluated on five public datasets. The results show that it has the best performance in almost all the settings compared to competitive methods such as ANTs, Elastix, Learning to Rank and Joint Label Fusion. Moreover, our registration method is more than 7 times as efficient as that of ANTs SyN, while our label transfer method is 18 times faster than Joint Label Fusion in CPU time. The results on the ADNI dataset demonstrate that our method is applicable to image pairs that require a significant transformation in registration. The performance on a composite dataset suggests that our method succeeds in a cross-modality manner. The results of this study show that the integrated 3D flow-based method is effective and efficient for brain structure segmentation. It also demonstrates the power of SIFT features, multi-atlas segmentation and classical machine learning algorithms for a medical image analysis task. The experimental results on public datasets show the proposed method's potential for general applicability in various brain structures and settings.

Advances in Mass Spectrometers for Flyby Space Missions for the Analysis of Biosignatures and Other Complex Molecules

Spacecraft flybys provide access to the chemical composition of the gaseous envelope of the planetary object. Typical relative encounter velocities range from km/s to tens of km/s in flybys. For speeds exceeding about 5 km/s, modern mass spectrometers analyzing the rapidly encountering gas suffer from intrinsic hypervelocity impact-induced fragmentation processes causing ambiguous results when analyzing complex molecules. In this case, instruments use an antechamber, inside which the incoming species collide many times with the chamber wall. These collisions cause the desired deceleration and thermalization of the gas molecules. However, these collisions also dissociate molecular bonds, thus fragmenting the molecules, and possibly forming new ones precluding scientists from inferring the actual chemical composition of the sampled gas. We developed a novel time-of-flight mass spectrometer that handles relative encounter velocities of up to 20 km/s omitting an antechamber and its related fragmentation. It analyzes the complete mass range of m/z 1 to 1000 at an instance. This innovation leads to unambiguous analysis of complex (organic) molecules. Applied to Enceladus, Europa or Io, it will provide reliable chemical composition datasets for exploration of the Solar System to determine its status, origin and evolution.

Research on Recognition of Quiet Period of Sandstone Acoustic Emission Based on Four Machine Learning Algorithms

Aiming at solving the problem that it is difficult to recognize the quiet period of acoustic emission in rocks, four machine learning algorithms were adopted to develop and improve the recognition method of the quiet period of acoustic emission. In the process of establishing the model, the time domain data of acoustic emission were standardized and processed by box diagram method, so as to clean the abnormal data and reduce the dimension, and the frequency domain data were denoised by wavelet four-layer transform and wavelet packet three-layer energy decomposition, and a group of 8 wavelet packet energy parameters were established as frequency domain characteristic parameters. Based on AE time domain data, frequency domain data, and composite data (time-frequency domain data sets), the grid search traversal parameter technique was used to obtain the optimal parameters of four machine learning models. The accuracy, precision, recall, and F 1 score were used to verify and evaluate the recognition performance of the models. The study results show that the recognition effects of the models are good, the model accuracy of the frequency domain data set is the lowest, and the model accuracy of the composite data set is the highest, with an accuracy of more than 90%. The kernel support vector machine model has the best performance, and its average precision is 0.87. The random forest (RF) model is the best model for recognizing quiet period of acoustic emission.

Development of a manufacturer-independent convolutional neural network for the automated identification of vertebral compression fractures in vertebral fracture assessment images using active learning.

Convolutional neural networks (CNNs) can identify vertebral compression fractures in GE vertebral fracture assessment (VFA) images with high balanced accuracy, but performance against Hologic VFAs is unknown. To obtain good classification performance, supervised machine learning requires balanced and labeled training data. Active learning is an iterative data annotation process with the ability to reduce the cost of labeling medical image data and reduce class imbalance. To train CNNs to identify vertebral fractures in Hologic VFAs using an active learning approach, and evaluate the ability of CNNs to generalize to both Hologic and GE VFA images. VFAs were obtained from the OsteoLaus Study (labeled Hologic Discovery A, n=2726), the Manitoba Bone Mineral Density Program (labeled GE Prodigy and iDXA, n=12,742), and the Canadian Longitudinal Study on Aging (CLSA, unlabeled Hologic Discovery A, n=17,190). Unlabeled CLSA VFAs were split into five equal-sized partitions (n=3438) and reviewed sequentially using active learning. Based on predicted fracture probability, 17.6% (n=3032) of the unlabeled VFAs were selected for expert review using the modified algorithm-based qualitative (mABQ) method. CNNs were simultaneously trained on Hologic, GE dual-energy and GE single-energy VFAs. Two ensemble CNNs were constructed using the maximum and mean predicted probability from six separately trained CNNs that differed due to stochastic variation. CNNs were evaluated against the OsteoLaus validation set (n=408) during the active learning process; ensemble performance was measured against the OsteoLaus test set (n=819). The baseline CNN, prior to active learning, achieved 55.0% sensitivity, 97.9% specificity, 57.9% positive predictive value (PPV), F1-score 56.4%. Through active learning, 2942 CLSA Hologic VFAs (492 fractures) were added to the training data-increasing the proportion of Hologic VFAs with fractures from 4.2% to 12.5%. With active learning, CNN performance improved to 80.0% sensitivity, 99.7% specificity, 94.1% PPV, F1-score 86.5%. The CNN maximum ensemble achieved 91.9% sensitivity (100% for grade 3 and 95.5% for grade 2 fractures), 99.0% specificity, 81.0% PPV, F1-score 86.1%. Simultaneously training on a composite dataset consisting of both Hologic and GE VFAs allowed for the development of a single manufacturer-independent CNN that generalized to both scanner types with good classification performance. Active learning can reduce class imbalance and produce an effective medical image classifier while only labeling a subset of available unlabeled image data-thereby reducing the time and cost required to train a machine learning model.

Rare earth elements plus yttrium (REY) in phosphorites from the Tébessa region (Eastern Algeria): Abundance, geochemical distribution through grain size fractions, and economic significance

Rare earth elements and yttrium (REY) have gained greater attention for being largely used in various high-tech applications and green energies. Recently, supply shortage and high demand on REY led to target secondary resources such as phosphorites. Algerian sedimentary phosphorites, which are located mainly in the Tébessa region, Eastern Saharan Atlas, are relatively enriched in REY according to recent whole-rock analyses. The aim of this study is to explore variation of REY contents and other geochemical features with respect to particle size fractions. In addition, the economic significance of the P-deposits was assessed using new indicators, such as the outlook coefficient of REY composition (Coutl) and the percentage of critical element in total ΣREY (REYdef). Three friable whole-rock samples from two P-deposits (Djebel El Kouif and Kef Essenoun) were sieved and retained fractions (f) of <45 μm, 45–125 μm, 125–250 μm, 250–500 μm and >500 μm were analyzed for their major and REY contents using ICP-MS techniques. Principal Component Analysis (PCA) was applied on centred log-transformation (clr) data, which are adapted to such compositional dataset, as well as Variance Analysis (one-way ANOVA) technique. The results show that the samples yield P2O5 grades ranging from 20 to 36 wt% and total REY contents in selected 15 samples vary in the considered grain size fractions as follows: f<45μm = 325 ppm–719 ppm; f45–125μm = 309 ppm–893 ppm; f125–250μm = 314 ppm–1029 ppm, f250–500μm = 354 ppm–809 ppm, f>500μm = 308 ppm–652 ppm. The one-way ANOVA reveals that there is no significant difference of REY concentrations between the grain size fractions. However, the REY grades strongly based on both the studied deposits and the stratigraphic position of the hosting samples (p-value ≤ 0.01); this is also confirmed by PCA. REY geochemical signatures (normalized REY distribution patterns and Ce, Eu, and Y anomalies) are similar in the grain size fractions from the same hosting samples indicating synchronized evolution of all rock components, where REY distribution is mostly controlled by the depositional environment, irrespective of grain size in the rock. The Coutl coefficient shows values between 1.89 and 6.85 and REYdef varies from 47.44 % to 64.90 %. However, the upper sub-layer of Kef Essenoun phosphorites that yielded the lowest ∑REY contents, shows markedly the highest Coutl (6.61–6.85) and REYdef (64.31–64.90), pointing to more promising source for extracting individual critical REY elements, even in low P-grade phosphorite ores.

Trace and Rare Earth Element Compositions of Lawsonite as a Chemical Tracer of Metamorphic Processes in Subduction Zones

AbstractLawsonite is a major host mineral of trace elements (TEs; e.g. REE, Sr, Pb, U, Th) and H2O in various rock types (metabasite, metasediment, metasomatite) over a wide range of depths in subduction zones. Consequently, the composition of lawsonite is a useful archive to track chemical exchanges that occurred during subduction and/or exhumation, as recorded in high-pressure/low-temperature (HP/LT) terranes. This study provides an extensive dataset of major element and TE compositions of lawsonite in HP/LT rocks from two mélanges (Franciscan/USA; Rio San Juan/Dominican Republic), two structurally coherent terranes (Tavşanlı/Turkey; Alpine Corsica/France), and the eclogite blocks of the Pinchi Lake/Canada complex. Bulk major and TE compositions were also determined for lawsonite-bearing host rocks to understand petrogenesis and assess compositional evolution. Most analyzed mélange and coherent-terrane metabasalts have normal mid-ocean ridge/back-arc basin basalt signatures and they preserve compositional evidence supporting interactions with (meta)sediment ± metagabbro/serpentinite (e.g. LILE/LREE enrichments; Ni/Cr enrichments). Most lawsonite grains analyzed are compositionally zoned in transition-metal elements (Fe, Ti, Cr), other TEs (e.g. Sr, Pb), and/or REE, with some grains showing compositional variations that correlate with zoning patterns (e.g. Ti-sector zoning, core-to-rim zoning in Fe, Cr-oscillatory zoning). Our results suggest that compositional variations in lawsonite formed in response to crystallographic control (in Ti-sector zoning), fluid–host rock interactions, modal changes in minerals, and/or element fractionation with coexisting minerals that compete for TEs (e.g. epidote, titanite). The Cr/V and Sr/Pb ratios of lawsonite are useful to track the compositional influence of serpentinite/metagabbro (high Cr/V) and quartz-rich (meta)sediment (low Sr/Pb). Therefore, lawsonite trace and rare earth element compositions effectively record element redistribution driven by metamorphic reactions and fluid–rock interactions that occurred in subduction systems.