Training Set Size Research Articles

A Strong Gravitational Lens Is Worth a Thousand Dark Matter Halos: Inference on Small-scale Structure Using Sequential Methods

Strong gravitational lenses are a singular probe of the Universe’s small-scale structure—they are sensitive to the gravitational effects of low-mass (<1010 M ⊙) halos even without a luminous counterpart. Recent strong-lensing analyses of dark matter structure rely on simulation-based inference (SBI). Modern SBI methods, which leverage neural networks as density estimators, have shown promise in extracting the halo-population signal. However, it is unclear whether the constraints from these models are limited by the methodology or the data. In this study, we introduce an accelerator-optimized simulation pipeline that can generate lens images with realistic subhalo populations in milliseconds. Leveraging this simulator, we identify the main limitation of our fiducial SBI analysis: training set size. We then adopt a sequential neural posterior estimation (SNPE) approach, allowing us to refine the training distribution to align with the observed data. Using only one-fifth as many mock Hubble Space Telescope images, SNPE matches the constraints on the low-mass halo population produced by our best nonsequential model. Our experiments suggest that an over 3 order-of-magnitude increase in training set size and GPU hours would be required to achieve an equivalent result without sequential methods. While the full potential of the existing lens sample remains to be explored, the notable improvement in constraining power enabled by our sequential approach highlights that current constraints are limited primarily by methodology and not the data itself. Moreover, our results emphasize the need to treat training set generation and model optimization as interconnected stages of any cosmological analysis using SBI.

MEF-AlloSite: an accurate and robust Multimodel Ensemble Feature selection for the Allosteric Site identification model

A crucial mechanism for controlling the actions of proteins is allostery. Allosteric modulators have the potential to provide many benefits compared to orthosteric ligands, such as increased selectivity and saturability of their effect. The identification of new allosteric sites presents prospects for the creation of innovative medications and enhances our comprehension of fundamental biological mechanisms. Allosteric sites are increasingly found in different protein families through various techniques, such as machine learning applications, which opens up possibilities for creating completely novel medications with a diverse variety of chemical structures. Machine learning methods, such as PASSer, exhibit limited efficacy in accurately finding allosteric binding sites when relying solely on 3D structural information.Scientific ContributionPrior to conducting feature selection for allosteric binding site identification, integration of supporting amino-acid–based information to 3D structural knowledge is advantageous. This approach can enhance performance by ensuring accuracy and robustness. Therefore, we have developed an accurate and robust model called Multimodel Ensemble Feature Selection for Allosteric Site Identification (MEF-AlloSite) after collecting 9460 relevant and diverse features from the literature to characterise pockets. The model employs an accurate and robust multimodal feature selection technique for the small training set size of only 90 proteins to improve predictive performance. This state-of-the-art technique increased the performance in allosteric binding site identification by selecting promising features from 9460 features. Also, the relationship between selected features and allosteric binding sites enlightened the understanding of complex allostery for proteins by analysing selected features. MEF-AlloSite and state-of-the-art allosteric site identification methods such as PASSer2.0 and PASSerRank have been tested on three test cases 51 times with a different split of the training set. The Student’s t test and Cohen’s D value have been used to evaluate the average precision and ROC AUC score distribution. On three test cases, most of the p-values (<0.05\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$< 0.05$$\\end{document}) and the majority of Cohen’s D values (>0.5\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$> 0.5$$\\end{document}) showed that MEF-AlloSite’s 1–6% higher mean of average precision and ROC AUC than state-of-the-art allosteric site identification methods are statistically significant.

Appearance-Based Gaze Estimation as a Benchmark for Eye Image Data Generation Methods

Data augmentation is commonly utilized to increase the size and diversity of training sets for deep learning tasks. In this study, we propose a novel application of an existing image generation approach in the domain of realistic eye images that leverages data collected from 40 subjects. This hybrid method combines the benefits of precise control over the image content provided by 3D rendering, while introducing the previously lacking photorealism and diversity into synthetic images through neural style transfer. We prove its general efficacy as a data augmentation tool for appearance-based gaze estimation when generated data are mixed with a sparse train set of real images. It improved the results for 39 out of 40 subjects, with an 11.22% mean and a 19.75% maximum decrease in gaze estimation error, achieving similar metrics for train and held-out subjects. We release our data repository of eye images with gaze labels used in this work for public access.

Scale matters: Large language models with billions (rather than millions) of parameters better match neural representations of natural language.

Recent research has used large language models (LLMs) to study the neural basis of naturalistic language processing in the human brain. LLMs have rapidly grown in complexity, leading to improved language processing capabilities. However, neuroscience researchers haven't kept up with the quick progress in LLM development. Here, we utilized several families of transformer-based LLMs to investigate the relationship between model size and their ability to capture linguistic information in the human brain. Crucially, a subset of LLMs were trained on a fixed training set, enabling us to dissociate model size from architecture and training set size. We used electrocorticography (ECoG) to measure neural activity in epilepsy patients while they listened to a 30-minute naturalistic audio story. We fit electrode-wise encoding models using contextual embeddings extracted from each hidden layer of the LLMs to predict word-level neural signals. In line with prior work, we found that larger LLMs better capture the structure of natural language and better predict neural activity. We also found a log-linear relationship where the encoding performance peaks in relatively earlier layers as model size increases. We also observed variations in the best-performing layer across different brain regions, corresponding to an organized language processing hierarchy.

Artificial intelligence for predicting arctic permafrost and active layer temperatures along the Alaskan North Slope

AbstractClimate change in Arctic regions poses a threat to permafrost stability, potentially leading to issues such as infrastructure damage, altered ecosystems, and increased greenhouse gas emissions. The challenge of monitoring permafrost temperatures and identifying critical environmental factors is accentuated by the scarcity of subsurface thermal data in remote Arctic locales. To alleviate this lack of soil thermal data we set out to combine in situ observed soil temperatures with MERRA-2 reanalysis data and Machine Learning (ML) to develop local and season-specific subsurface soil temperature predictions in Alaska. First, reanalysis features are selected based on surface energy budget physics. Coupled with Julian calendar dates, reanalysis features are preprocessed and then fed to the ML prediction model. We conducted a series of computational experiments and tested the results against performance benchmarks to determine the optimal prediction models, length of look-back period of model inputs, and the training set size. Six conventional ML models (e.g., GBDT, RF, SVR) and five statistical baselines (e.g. ARIMA) using in situ time series of soil temperature data ranging from 0 - 1 m depth are considered for the prediction task. The models are trained using in situ soil temperatures at depths between 0 and 1 m, spanning 16+ years, from field sites at Deadhorse and Toolik Lake, Alaska. All prediction performances are assessed using root mean squared error (RMSE) and mean absolute error (MAE). Results show that locally trained ML models can estimate shallow soil temperatures with an average error of $$RMSE=1.308^{\circ }$$ R M S E = 1 . 308 ∘ C.

Improving machine-learning models in materials science through large datasets

The accuracy of a machine learning model is limited by the quality and quantity of the data available for its training and validation. This problem is particularly challenging in materials science, where large, high-quality, and consistent datasets are scarce. Here we present alexandria, an open database of more than 5 million density-functional theory calculations for periodic three-, two-, and one-dimensional compounds. We use this data to train machine learning models to reproduce seven different properties using both composition-based models and crystal-graph neural networks. In the majority of cases, the error of the models decreases monotonically with the training data, although some graph networks seem to saturate for large training set sizes. Differences in the training can be correlated with the statistical distribution of the different properties. We also observe that graph-networks, that have access to detailed geometrical information, yield in general more accurate models than simple composition-based methods. Finally, we assess several universal machine learning interatomic potentials. Crystal geometries optimised with these force fields are very high quality, but unfortunately the accuracy of the energies is still lacking. Furthermore, we observe some instabilities for regions of chemical space that are undersampled in the training sets used for these models. This study highlights the potential of large-scale, high-quality datasets to improve machine learning models in materials science.

Will Transformers change gastrointestinal endoscopic image analysis? A comparative analysis between CNNs and Transformers, in terms of performance, robustness and generalization

Gastrointestinal endoscopic image analysis presents significant challenges, such as considerable variations in quality due to the challenging in-body imaging environment, the often-subtle nature of abnormalities with low interobserver agreement, and the need for real-time processing. These challenges pose strong requirements on the performance, generalization, robustness and complexity of deep learning-based techniques in such safety–critical applications. While Convolutional Neural Networks (CNNs) have been the go-to architecture for endoscopic image analysis, recent successes of the Transformer architecture in computer vision raise the possibility to update this conclusion. To this end, we evaluate and compare clinically relevant performance, generalization and robustness of state-of-the-art CNNs and Transformers for neoplasia detection in Barrett’s esophagus. We have trained and validated several top-performing CNNs and Transformers on a total of 10,208 images (2,079 patients), and tested on a total of 7,118 images (998 patients) across multiple test sets, including a high-quality test set, two internal and two external generalization test sets, and a robustness test set. Furthermore, to expand the scope of the study, we have conducted the performance and robustness comparisons for colonic polyp segmentation (Kvasir-SEG) and angiodysplasia detection (Giana). The results obtained for featured models across a wide range of training set sizes demonstrate that Transformers achieve comparable performance as CNNs on various applications, show comparable or slightly improved generalization capabilities and offer equally strong resilience and robustness against common image corruptions and perturbations. These findings confirm the viability of the Transformer architecture, particularly suited to the dynamic nature of endoscopic video analysis, characterized by fluctuating image quality, appearance and equipment configurations in transition from hospital to hospital. The code is made publicly available at: https://github.com/BONS-AI-VCA-AMC/Endoscopy-CNNs-vs-Transformers.

Improved Monitoring and Classification of Engine Oil Condition through Two Machine Learning Techniques

&lt;div&gt;This study explores the effectiveness of two machine learning models, namely multilayer perceptron neural networks (MLP-NN) and adaptive neuro-fuzzy inference systems (ANFIS), in advancing maintenance management based on engine oil analysis. Data obtained from a Mercedes Benz 2628 diesel engine were utilized to both train and assess the MLP-NN and ANFIS models. Six indices—Fe, Pb, Al, Cr, Si, and PQ—were employed as inputs to predict and classify engine conditions. Remarkably, both models exhibited high accuracy, achieving an average precision of 94%. While the radial basis function (RBF) model, as presented in a referenced article, surpassed ANFIS, this comparison underscored the transformative potential of artificial intelligence (AI) tools in the realm of maintenance management. Serving as a proof-of-concept for AI applications in maintenance management, this study encourages industry stakeholders to explore analogous methodologies.&lt;/div&gt; &lt;section&gt; &lt;h2&gt;Highlights&lt;/h2&gt; &lt;div&gt; &lt;ul&gt; &lt;li&gt; &lt;div&gt;Two machine learning models, multilayer perceptron neural networks (MLP-NN) and adaptive neuro-fuzzy inference systems (ANFIS), were employed to predict and classify the performance condition of diesel engines.&lt;/div&gt; &lt;/li&gt; &lt;li&gt; &lt;div&gt;Among various training algorithms, Levenberg–Marquardt and the Bayesian regularization demonstrated superior classification accuracy, achieving a 95%–96% range.&lt;/div&gt; &lt;/li&gt; &lt;li&gt; &lt;div&gt;To assess the generalizability of MLP-NN and ANFIS, the training set size was varied from 90% to 10%.&lt;/div&gt; &lt;/li&gt; &lt;li&gt; &lt;div&gt;The ANFIS model exhibited greater stability than MLP-NN, with a 50% higher performance.&lt;/div&gt; &lt;/li&gt; &lt;/ul&gt; &lt;/div&gt; &lt;/section&gt; &lt;section&gt; &lt;h2&gt;Graphical Abstract&lt;/h2&gt; &lt;div&gt; &lt;img/&gt; &lt;/div&gt; &lt;/section&gt;

Performance comparison of machine learning algorithms for the estimation of blood pressure using photoplethysmography

This paper deals with an in-depth performance analysis on the estimation of Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP) by using features from the photoplethysmography (PPG) signal enhanced using the Maximal Overlap Discrete Wavelet Transform (MODWT), to train many machine learning (ML) regression models, including eXtreme Gradient Boost (XGBoost). The impact of different features selections methods, ML methods and training set sizes has been analyzed. One result is on the achievable improvements using features extracted from MODWT enhanced PPG signals. The most significant features have been selected using three different algorithms, namely RReliefF, Minimum Redundancy Maximum Relevance (MRMR) and Correlation-based Feature Selection (CFS). This comparison has been critical to underline that the exploitation of the new features allows to improve SBP and DBP estimation. Moreover, the authors have trained several ML algorithms to provide a comparison of their accuracy and training time, showing the Pareto frontier. RReliefF and MRMR selections algorithms, and several ML algorithms such as XGBoost, Gaussian Process Regression(GPR) and Ensemble stood out for their performance, with a different compromise between prediction error and training time. In addition, a further result has been obtained by varying the dimension of the dataset to understand the impact on Root Mean Square Error (RMSE) for models that have shown better performance, giving an empirical relationship on achievable RMSE as a function of training set size. From that relationship it has been extrapolated an upper boundary of the set size over which no further RMSE improvements are expected.

A short video sentiment analysis model based on multimodal feature fusion

With the development of the internet, the number of short video platform users has increased quickly. People's social entertainment mode has gradually changed from text to short video, generating many multimodal data. Therefore, traditional single-modal sentiment analysis can no longer fully adapt to multimodal data. To address this issue, this study proposes a short video sentiment analysis model based on multimodal feature fusion. This model analyzes the text, speech, and visual content in the video. Meanwhile, the information of the three modalities is integrated through a multi-head attention mechanism to analyze and classify emotions. The experimental results showed that when the training set size was 500, the recognition accuracy of the multimodal sentiment analysis model based on modal contribution recognition and multi-task learning was 0.96. The F1 score was 98, and the average absolute error value was 0.21. When the validation set size was 400, the recognition time of the multimodal sentiment analysis model based on modal contribution recognition and multi-task learning was 2.1 s. When the iterations were 60, the recognition time of the multimodal sentiment analysis model based on modal contribution recognition and multi-task learning was 0.9 s. The experimental results show that the proposed multimodal sentiment analysis model based on modal contribution recognition and multi-task learning has good model performance and can accurately identify emotions in short videos.

Self-supervised contrastive learning improves machine learning discrimination of full thickness macular holes from epiretinal membranes in retinal OCT scans.

There is a growing interest in using computer-assisted models for the detection of macular conditions using optical coherence tomography (OCT) data. As the quantity of clinical scan data of specific conditions is limited, these models are typically developed by fine-tuning a generalized network to classify specific macular conditions of interest. Full thickness macular holes (FTMH) present a condition requiring urgent surgical repair to prevent vision loss. Other works on automated FTMH classification have tended to use supervised ImageNet pre-trained networks with good results but leave room for improvement. In this paper, we develop a model for FTMH classification using OCT B-scans around the central foveal region to pre-train a naïve network using contrastive self-supervised learning. We found that self-supervised pre-trained networks outperform ImageNet pre-trained networks despite a small training set size (284 eyes total, 51 FTMH+ eyes, 3 B-scans from each eye). On three replicate data splits, 3D spatial contrast pre-training yields a model with an average F1-score of 1.0 on holdout data (50 eyes total, 10 FTMH+), compared to an average F1-score of 0.831 for FTMH detection by ImageNet pre-trained models. These results demonstrate that even limited data may be applied toward self-supervised pre-training to substantially improve performance for FTMH classification, indicating applicability toward other OCT-based problems.

Anomaly detection combining bidirectional gated recurrent unit and autoencoder in the context of e-commerce

E-commerce platforms store a large amount of user personal information, transaction data, and financial information, which have extremely high value for hackers and criminals. Therefore, protecting the security of e-commerce platforms is particularly important, and intrusion detection is a technical means used to discover and respond to possible security threats and attacks. But with the development of Internet technology, there are more and more types of intrusion attacks and more sophisticated means. Traditional intrusion detection systems are difficult to cope with. This study proposes an anomaly detection model based on bidirectional gated loop units and autoencoders. The model learns HTTP text data, trains the model, and uses bidirectional gated loop units to convert text sequences from characters to numbers. The experimental results show that when the training set size is 1000, the false alarm rates of Analytic Hierarchy Process, Support Vector Machine, Long Short Term Recurrent Memory Network, and Improved end-to-end algorithm models are 0.30, 0.27, 0.23, and 0.10, respectively. The loss function values are 0.35, 0.28, 0.17, and 0.13, respectively. The F1 values are 0.78, 0.88, 0.91, and 0.99, and the accuracy rates are 0.88, 0.91, 0.95, and 0.99, respectively. The research results indicate that the proposed method model has excellent performance.

Multi-trait multi-environment genomic prediction of preliminary yield trial in pulse crop.

Phenotypic selection of complex traits such as seed yield and protein in the preliminary yield trial (PYT) is often constrained by limited seed availability, resulting in trials with few environments and minimal to no replications. Multi-trait multi-environment enabled genomic prediction (MTME-GP) offers a valuable alternative to predict missing phenotypes of selection candidates for multiple traits and diverse environments. In this study, we assessed the efficiency of MTME-GP for improving seed protein and seed yield in field pea, the top two breeding targets but highly antagonistic traits in pulse crop. We utilized a set of 300 selection candidates in the PYT that virtually represented all possible families of the North Dakota State University field pea breeding program. Selection candidates were evaluated in three diverse, contrasting environments, as indicated by a range of heritability. Using whole- and split-environment cross validation schemes, MTME-GP had higher predictive ability than a standard additive G-BLUP model. Integrating a range of overlapping genotypes in between environments showed improvement on the predictive ability of the MTME-GP model but tends to plateau at 50%-80% training set size. Regardless of the cross-validation scheme, accuracy was among the lowest in stressed environments, presumably due to low heritability for seed protein and yield. This study provided insights into the potential of MTME-GP in a public pulse crop breeding program. The MTME-GP framework can be further improved with more testing environments and integration of additional orthogonal information in the early stages of the breeding pipeline.

Deep learning-based automated liver contouring using a small sample of radiotherapy planning computed tomography images

IntroductionNo study has yet investigated the minimum amount of data required for deep learning-based liver contouring. Therefore, this study aimed to investigate the feasibility of automated liver contouring using limited data. MethodsRadiotherapy planning Computed tomography (CT) images were subjected to various preprocessing methods, such as denoising and windowing. Segmentation was conducted using the modified Attention U-Net and Residual U-Net networks. Two different modified networks were trained separately for different training sizes. For each architecture, the model trained with the training set size that achieved the highest dice similarity coefficient (DSC) score was selected for further evaluation. Two unseen external datasets with different distributions from the training set were also used to examine the generalizability of the proposed method. ResultsThe modified Residual U-Net and Attention U-Net networks achieved average DSCs of 97.62% and 96.48%, respectively, on the test set, using 62 training cases. The average Hausdorff distances (AHDs) for the modified Residual U-Net and Attention U-Net networks were 0.57 mm and 0.71 mm, respectively. Also, the modified Residual U-Net and Attention U-Net networks were tested on two unseen external datasets, achieving DSCs of 95.35% and 95.82% for data from another center and 95.16% and 94.93% for the AbdomenCT-1K dataset, respectively. ConclusionThis study demonstrates that deep learning models can accurately segment livers using a small training set. The method, utilizing simple preprocessing and modified network architectures, shows strong performance on unseen datasets, indicating its generalizability. Implications for practiceThis promising result suggests its potential for automated liver contouring in radiotherapy planning.

A deep belief network-based energy consumption prediction model for water source heat pump system

To achieve optimal energy efficiency in buildings, accurately forecasting the energy consumption of air conditioning systems is crucial. This study develops an energy consumption prediction model based on a deep belief network, which is constructed according to the principles of a restricted Boltzmann machine. Actual experimental data from a water source heat pump system are collected, and feature variables are selected. The study discusses the impact of model parameters and training set sizes on the performance of energy consumption prediction model. Additionally, the trend in model prediction performance is analyzed through parameter adjustments. The results show that the coefficient of determination (R2) for the optimized model has increased to 0.585. The mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) have been reduced to 6.311, 2.512, and 1.625, respectively. The deep belief network energy consumption prediction model outperforms other common machine learning models for water source heat pump systems.

Application of a Machine Learning-Based Classification Approach for Developing Host Protein Diagnostic Models for Infectious Disease

In recent years, infectious disease diagnosis has increasingly turned to host-centered approaches as a complement to pathogen-directed ones. The former, however, typically requires the interpretation of complex multiple biomarker datasets to arrive at an informative diagnostic outcome. This report describes a machine learning (ML)-based classification workflow that is intended as a template for researchers seeking to apply ML approaches for developing host-based infectious disease biomarker classifiers. As an example, we built a classification model that could accurately distinguish between three disease etiology classes: bacterial, viral, and normal in human sera using host protein biomarkers of known diagnostic utility. After collecting protein data from known disease samples, we trained a series of increasingly complex Auto-ML models until arriving at an optimized classifier that could differentiate viral, bacterial, and non-disease samples. Even when limited to a relatively small training set size, the model had robust diagnostic characteristics and performed well when faced with a blinded sample set. We present here a flexible approach for applying an Auto-ML-based workflow for the identification of host biomarker classifiers with diagnostic utility for infectious disease, and which can readily be adapted for multiple biomarker classes and disease states.

A meta-analysis on the effects of marker coverage, status number, and size of training set on predictive accuracy and heritability estimates from genomic selection in tree breeding

Genomic selection (GS) is increasingly used in tree breeding because of the possibility to hasten breeding cycles, increase selection intensity or facilitate multi-trait selection, and to obtain less biased estimates of quantitative genetic parameters such as heritability. However, tree breeders are aiming to obtain accurate estimates of such parameters and breeding values while optimizing sampling and genotyping costs. We conducted a metadata analysis of results from 28 GS studies totalling 115 study-traits. We found that heritability estimates obtained using DNA marker-based information for a variety of traits and species were not significantly related to variation in the total number of markers ranging from about 1500 to 116 000, nor by the marker density, ranging from about 1 to 60 markers/centimorgan, nor by the status number of the breeding populations ranging from about 10 to 620, nor by the size of the training set ranging from 236 to 2458. However, the predictive accuracy of breeding values was generally higher when the status number of the breeding population was smaller, which was expected given the higher level of relatedness in small breeding populations, and the increased ability of a given number of markers to trace the long-range linkage disequilibrium in such conditions. According to expectations, the predictive accuracy also increased with the size of the training set used to build marker-based models. Genotyping arrays with a few to many thousand markers exist for several tree species and with the actual costs, GS could thus be efficiently implemented in many more tree breeding programs, delivering less biased genetic parameters and more accurate estimates of breeding values.

Layer-by-layer reconstruction of fatigue damages in composites from thermal images by a Residual U-Net

In this paper a deep learning model is used to fully reconstruct the 3D distribution of arbitrarily shaped subsurface fatigue damages in a fiber/epoxy composite from synthetic thermal surface images. Synthetic thermal surface images (TIs) of self-heating damage hotspots are produced by thermal finite element analysis which are consequently used to train a Residual U-Net based on recent architectures designed for image segmentation. Different augmentation techniques are employed to mitigate the computational cost of generating training data through thermal finite element analysis. The Residual U-Net model accurately reconstructed – layer by layer – the ground truths and thereby enabled the quantitative assessment of location, size, shape, depth and gradient of an internal fatigue damage distribution. Moreover, the Residual U-Net achieved good predictions for comparatively small training set sizes.

A Hybrid Neural Network-Based Fast Financial Fraud Detection Model

With the increasing number of financial transactions, financial fraud has become increasingly serious for financial institutions and the public. The core idea of this model is to integrate multiple neural network structures and utilize their respective advantages to improve the performance of fraud detection. Firstly, we employed the convolutional neural network with interpretable blocks (CNNIB) convolutional neural network (CNN) to extract key features from the data to capture patterns and patterns in fraud cases. Secondly, we introduced the autoencoder generative adversarial network (AE-GAN) adversarial network to perform feature analysis on sequence data to capture temporal features in transaction sequences. Finally, we used differential detection for classification to determine whether transactions were fraudulent. An independent detection module was established to accelerate the recognition of financial fraud, and parameter indicators were optimized. Finally, a hybrid neural network model was established. The experimental results indicate that our model has achieved significant results in quickly detecting financial fraud; compared with traditional single neural network models, hybrid neural network models have significant improvements in accuracy and efficiency. In addition, we conducted in-depth analysis of the model and revealed its performance stability under different training set sizes and data distributions. Our research findings provide an effective tool for financial institutions to quickly identify financial fraud.

A Comparison of CT-Based Pancreatic Segmentation Deep Learning Models

Rationale and ObjectivesPancreas segmentation accuracy at CT is critical for the identification of pancreatic pathologies and is essential for the development of imaging biomarkers. Our objective was to benchmark the performance of five high-performing pancreas segmentation models across multiple metrics stratified by scan and patient/pancreatic characteristics that may affect segmentation performance. Materials and MethodsIn this retrospective study, PubMed and ArXiv searches were conducted to identify pancreas segmentation models which were then evaluated on a set of annotated imaging datasets. Results (Dice score, Hausdorff distance [HD], average surface distance [ASD]) were stratified by contrast status and quartiles of peri-pancreatic attenuation (5mm region around pancreas). Multivariate regression was performed to identify imaging characteristics and biomarkers (n=9) that were significantly associated with Dice score. ResultsFive pancreas segmentation models were identified: Abdomen Atlas [AAUNet, AASwin, trained on 8448 scans], TotalSegmentator [TS, 1204 scans], nnUNetv1 [MSD-nnUNet, 282 scans], and a U-Net based model for predicting diabetes [DM-UNet, 427 scans]. These were evaluated on 352 CT scans (30 females, 25 males, age 58±7 years old [±1SD], 327 age unknown) from 2000-2023. Overall, TS, AAUNet, and AASwin were the best performers, Dice=80±11%, 79±16%, and 77±18% respectively (pairwise Sidak test not-significantly different). AASwin and MSD-nnUNet performed worse (for all metrics) on non-contrast scans (vs contrast, P<.001). The worst performer was DM-UNet (Dice=67±16%). All algorithms except TS showed lower Dice scores with increasing peri-pancreatic attenuation (P<.01). Multivariate regression showed non-contrast scans, (P<.001; MSD-nnUNet), smaller pancreatic length (P=.005, MSD-nnUNet), and height (P=.003, DM-UNet) were associated with lower Dice scores. ConclusionThe convolutional neural network-based models trained on a diverse set of scans performed best (TS, AAUnet, and AASwin). TS performed equivalently to AAUnet and AASwin with only 13% of the training set size (8488 vs 1204 scans). Though trained on the same dataset, a transformer network (AASwin) had poorer performance on non-contrast scans whereas its convolutional network counterpart (AAUNet) did not. This study highlights how aggregate assessment metrics of pancreatic segmentation algorithms seen in other literature are not enough to capture differential performance across common patient and scanning characteristics in clinical populations.