Small Dataset Size Research Articles

Designing the next generation of polymers with machine learning and physics-based models

Abstract The development of next-generation polymers necessitates optimizing several key properties simultaneously, a task that is expensive and infeasible using traditional trial-and-error experimental approaches. A promising alternative is employing a combination of machine learning and physics-based tools to rapidly screen the polymer design space and provide suggestions of new polymers that meet the critical properties required for industrial applications. In this study, we introduce a comprehensive workflow that utilizes machine learning and molecular modeling approaches to design new polymers with the focus on improving five polymer properties: (1) glass transition temperature, (2) dielectric constant, (3) refractive index, (4) stress optic coefficient, and (5) linear coefficient of thermal expansion. Using a small dataset ( < 200 unique polymers), we developed quantitative structure-property relationships (QSPRs) models to accurately predict the experimental polymer properties for both homo- and co-polymer systems. We tested several ML algorithms and identified the best models for predicting these polymer properties, achieving test set R 2 greater than 0.77 across all properties. We then explored new polymers by creating a library of over ∼10 000 homopolymers using R-group enumeration tools and applied the trained QSPR models to rapidly predict the five polymer properties. The predictions of QSPR models were used to create a multi-parameter optimization score, which helped downselect the large polymer space to ∼10 promising candidates. The properties of these selected polymer candidates were subsequently validated with classical molecular dynamics simulations and density functional theory, revealing a strong correlation with the QSPR model predictions. Finally, one of the top candidates was validated by experiments, which showed good agreement against QSPR and physics-based models. Our workflow underscores the power of combining data-driven and theoretical methods in the polymer design process given a small dataset size, offering a valuable resource for experimentalists looking to leverage computer-aided strategies in materials innovation.

Early detection of age-related macular degeneration using Vision Transformer-based Architectures – A comparative study with offline metrics and data augmenting

Age-related macular degeneration (AMD) is one of the leading causes of vision loss in elderly adults around the world and is among the main visual impairments in Mexico. The difficulty of diagnosing AMD in its early stages motivates the use of advanced deep-learning methods that offer significant potential to improve diagnostic accuracy in retinal image analysis. In recent years, Transformer architectures for computer vision, such as Vision Transformer (ViT), Swin Transformer and BERT Pre-training of Image Transformers (BEiT) have provided a novel perspective for image analysis. This study presents a comparative analysis of these architectures, applied to AMD detection, focusing on each model's capability to classify the early stages of the disease. Although the small size of medical image datasets represented a challenge, our results suggest that ViT-based architectures and their derivatives achieve significant performance in AMD detection. BEiT is particularly notable for its consistently superior performance.

CadastreVision: A benchmark dataset for cadastral boundary delineation from multi-resolution earth observation images

Approximately 70%–75% of people worldwide have no formally registered land rights. Fit-For-Purpose Land Administration was introduced to address this problem and focuses on delineating visible cadastral boundaries from earth observation imagery. Recent studies have shown the potential of deep learning models to extract these visible cadastral boundaries automatically. However, studies are limited by the small size and geographical coverage of available datasets and by the lack of information about which cadastral boundaries are visible, i.e., associated with a physical object boundary. To overcome these problems, we present CadastreVision, a benchmark dataset containing cadastral reference data and corresponding multi-resolution earth observation imagery from The Netherlands, with a spatial resolution ranging from 0.1 m to 10 m. The ratio between visible and non-visible cadastral boundaries is essential to evaluate the potential automation level in cadastral boundary extraction from earth observation images and interpret results obtained by deep learning models. We investigate this ratio using a novel analysis pipeline that overlays cadastral reference data with visible topographic object boundaries. Our results show that approximately 72% of the total length of cadastral boundaries in The Netherlands are visible. CadastreVision will enable new developments in cadastral boundary delineation and future endeavours to investigate knowledge transfer to data-scarce areas. Our data and code is available at https://github.com/jeroengrift/cadastrevision.

SChemNET: a deep learning framework for predicting small molecules targeting microRNA function

MicroRNAs (miRNAs) have been implicated in human disorders, from cancers to infectious diseases. Targeting miRNAs or their target genes with small molecules offers opportunities to modulate dysregulated cellular processes linked to diseases. Yet, predicting small molecules associated with miRNAs remains challenging due to the small size of small molecule-miRNA datasets. Herein, we develop a generalized deep learning framework, sChemNET, for predicting small molecules affecting miRNA bioactivity based on chemical structure and sequence information. sChemNET overcomes the limitation of sparse chemical information by an objective function that allows the neural network to learn chemical space from a large body of chemical structures yet unknown to affect miRNAs. We experimentally validated small molecules predicted to act on miR-451 or its targets and tested their role in erythrocyte maturation during zebrafish embryogenesis. We also tested small molecules targeting the miR-181 network and other miRNAs using in-vitro and in-vivo experiments. We demonstrate that our machine-learning framework can predict bioactive small molecules targeting miRNAs or their targets in humans and other mammalian organisms.

Chemoinformatics for corrosion science: Data-driven modeling of corrosion inhibition by organic molecules.

This paper reviews the application of machine learning to the inhibition of corrosion by organic molecules. The methodologies considered include quantitative structure-property relationships (QSPR) and related data-driven approaches. The characteristic features of their key components are considered as applied to corrosion inhibition, including datasets, response properties, molecular descriptors, machine learning methods, and structure-property models. It is shown that the most important factors determining their choice and application features are: (1) the small or very small size of datasets, (2) the mechanism of corrosion inhibition associated with the adsorption of inhibitor molecules on the metal surface, and (3) multifactorial conditioning and noisiness of response property. On this basis, the application of machine learning to the inhibition of corrosion of materials based on iron, aluminum, and magnesium is considered. The main trends in the development of QSPR and related data-driven modeling of corrosion inhibition are discussed, the shortcomings and common errors are considered, and the prospects for their further development are outlined.

Forecasting the equity premium: can machine learning beat the historical average?

We empirically predict the equity premium with the selected machine learning methods in Gu et al. (Empirical asset pricing via machine learning. Rev. Financ. Stud., 2020, 33(5), 2223–2273). We also consider four additional popular machine learning methods (ridge regression, support vector regression, k-nearest neighbors, and extreme gradient boosted trees) and their combination method. Using a dataset of both macroeconomic and technical predictors, we find that despite showcasing strong in-sample forecasting abilities, particularly with tree-based models, the out-of-sample results support Welch and Goyal (A comprehensive look at the empirical performance of equity premium prediction. Rev. Financ. Stud., 2008, 21(4), 1455–1508) that the competing forecasting models generally fail to outperform the historical average benchmark. We attribute this failure to the small dataset size and the low signal-to-noise ratio inherent in equity premium prediction. Our variable importance analysis further identifies three bond interest rate-related variables as the most dominant predictors for the equity premium. The economic value from a market timing perspective highlights that the historical average benchmark strategy generates the highest average return of 25.63% and the best Sharpe ratio of 0.8. Finally, our findings are robust across a variety of settings.

Ultrasonic Weld Quality Inspection Involving Strength Prediction and Defect Detection in Data-Constrained Training Environments.

Welding is an extensively used technique in manufacturing, and as for every other process, there is the potential for defects in the weld joint that could be catastrophic to the manufactured products. Different welding processes use different parameter settings, which greatly impact the quality of the final welded products. The focus of research in weld defect detection is to develop a non-destructive testing method for weld quality assessment based on observing the weld with an RGB camera. Deep learning techniques have been widely used in the domain of weld defect detection in recent times, but the majority of them use, for example, X-ray images. An RGB image-based solution is attractive, as RGB cameras are comparatively inexpensive compared to X-ray image solutions. However, the number of publicly available RGB image datasets for weld defect detection is comparatively lower than that of X-ray image datasets. This work achieves a complete weld quality assessment involving lap shear strength prediction and visual weld defect detection from an extremely limited dataset. First, a multimodal dataset is generated by the fusion of image data features extracted using a convolutional autoencoder (CAE) designed in this experiment and input parameter settings data. The fusion of the dataset reduced lap shear strength (LSS) prediction errors by 34% compared to prediction errors using only input parameter settings data. This is a promising result, considering the extremely small dataset size. This work also achieves visual weld defect detection on the same limited dataset with the help of an ultrasonic weld defect dataset generated using offline and online data augmentation. The weld defect detection achieves an accuracy of 74%, again a promising result that meets standard requirements. The combination of lap shear strength prediction and visual defect detection leads to a complete inspection to avoid premature failure of the ultrasonic weld joints. The weld defect detection was compared against the publicly available image dataset for surface defect detection.

Ontology-Based Data Collection for a Hybrid Outbreak Detection Method Using Social Media.

Given the persistent global challenge presented by rapidly spreading diseases, as evidenced notably by the widespread impact of the COVID-19 pandemic on both human health and economies worldwide, the necessity of developing effective infectious disease prediction models has become of utmost importance. In this context, the utilization of online social media platforms as valuable tools in healthcare settings has gained prominence, offering direct avenues for disseminating critical health information to the public in a timely and accessible manner. Propelled by the ubiquitous accessibility of the internet through computers and mobile devices, these platforms promise to revolutionize traditional detection methods, providing more immediate and reliable epidemiological insights. Leveraging this paradigm shift, our proposed framework harnesses Twitter data associated with infectious disease symptoms, employing ontology to identify and curate relevant tweets. Central to our methodology is a hybrid model that integrates XGBoost and Bidirectional Long Short-Term Memory (BiLSTM) architectures. The integration of XGBoost addresses the challenge of handling small dataset sizes, inherent during outbreaks due to limited time series data. XGBoost serves as a cornerstone for minimizing the loss function and identifying optimal features from our multivariate time series data. Subsequently, the combined dataset, comprising original features and predicted values by XGBoost, is channeled into the BiLSTM for further processing. Through extensive experimentation with a dataset spanning multiple infectious disease outbreaks, our hybrid model demonstrates superior predictive performance compared to state-of-the-art and baseline models. By enhancing forecasting accuracy and outbreak tracking capabilities, our model offers promising prospects for assisting health authorities in mitigating fatalities and proactively preparing for potential outbreaks.

Forest Fire Hazard Forecasting Based on Google Earth Engine Open Satellite Data

Introduction. Forest fires cause significant damage to both the natural fund and the national economy. In recent years, their harmful influence has increased due to global warming, and in Ukraine also due to the armed conflict. Thus, traditional methods of patrolling (ground and air) are not only costly, but also dangerous due to mines and possible shelling of border areas. Therefore, the role of surveillance using satellite systems is increasing. Space monitoring is more efficient and covers a larger area of ??the Earth's surface. Another important advantage is open access to information. The purpose of the paper is to build a mathematical model for determining fire danger based on climatic and biophysical satellite data for the forests of Ukraine, as well as a similar climatic zone with the possibility of further scaling to other climatic regions and types of vegetation cover. To adhere to the principles of open science, Google Earth Engine (GEE), a cloud-based platform that provides open access to dynamic collections of pre-processed Earth remote sensing results, was chosen. Results. Climatic and biophysical data for the forests of Ukraine for the years 2017-2020 were collected using the tools of the Python library for working with GEE data. Further, the obtained data were processed by two methods: linear (PCA) and non-linear (UMAP) in order to obtain statistically independent attributes. Both obtained datasets were subjected to statistical processing using the Bayesian method. Finally, for each point on the map for which information was collected, an indicator was calculated that predicted fire danger if the obtained value was greater than 1 and its absence in the opposite case. The resulting model showed its efficiency on training data. On the test dataset (data on Polish forests for the same period), the results turned out to be worse, in particular, the model using PCA did not predict absence of fire danger, and the model using UMAP generally showed lower performance. This can be due to both the imperfection of the model and the small size of the test dataset or factors unrelated to natural processes (in particular the human factor). Conclusions. An approach to forest fires forecasting based on satellite data is proposed. The obtained results indicate that the model is already effective at this stage, although machine learning methods have not yet been applied. However, it needs further improvement, so work on the model will be continued. Along with improving the quality of forecasts, attention will be paid to the geographic expansion and the creation of a web application. Keywords: fire danger in forests, satellite data, correlation, PCA, UMAP, hard-to-reach areas.

Childhood growth of singletons conceived following intracytoplasmic sperm injection - irrelevance of gonadotropin stimulation.

In conventional, gonadotropin stimulated, in vitro fertilization or intracytoplasmic sperm injection (c-IVF/ICSI) growth and development of multiple follicles is induced by gonadotropins, combined with gonadotropin-releasing hormone agonist or antagonist. In recent studies, singletons conceived after c-IVF/ICSI cycles had lower birth weight not only than spontaneously conceived children but also children born after unstimulated natural IVF/ICSI cycles (NC-IVF/ICSI). Lower birth weight is associated with a catch-up growth within the first years of life. Following the Barker hypothesis accelerated growth has been associated with a higher risk of cardiovascular diseases later in life. The aim of the study is to assess, if children conceived with NC-IVF/ICSI have a higher birthweight and therefore do not show a catch-up growth within the first two years. Therefore, we assume that children born after NC-IVF/ICSI have a better long-term cardiometabolic risk profile. Whether the weight- and height gain is comparable to spontaneously conceived children is unknown, since to our knowledge we are the first study to investigate the longitudinal growth of children born after unstimulated natural cycle ICSI (NC-ICSI). We conducted a single-center, prospective cohort study (2010-2017) including children (n = 139) born after NC-ICSI or c-ICSI treatment. Growth parameters up to 24 months were collected. Standard deviation scores based on growth references were calculated. The study included 98 children in the NC-ICSI and 41 children in the c-ICSI group. The median birth weight in NC-ICSI children was 3.4 kg [0.1 standard deviation score (SDS)] compared to 3.3 kg (-0.3 SDS) in c-ICSI children (p = 0.61). Median length at birth was 50 cm in both groups (NC-ICSI (-0.5 SDS), c-ICSI children (-0.8 SDS), p = 0.48). At age 24 months, median weight in NC-ICSI children was 12.2 kg (0.3 SDS) versus 12.2 kg (0.2 SDS) in c-ICSI children (p = 0.82) and median length 87.5 cm (0.1 SDS) versus 88.0 cm (0.4 SDS) (p = 0.43). We found no difference in growth between children conceived after stimulated and unstimulated ICSI. Growth parameters of both treatment groups did not differ from Swiss national growth references (N = 8500). One of the main limitations of our study was the small sample size (N = 139) of complete data sets over time and the high drop-out rate of 49% (68/139). Nevertheless, with the increasing number of children born after IVF/ICSI every year it is of immense importance to search for possibilities to reduce their long-term cardiometabolic risk and we want our data to contribute to this discussion.

PharmaBench: Enhancing ADMET benchmarks with large language models

Accurately predicting ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties early in drug development is essential for selecting compounds with optimal pharmacokinetics and minimal toxicity. Existing ADMET-related benchmark sets are limited in utility due to their small dataset sizes and the lack of representation of compounds used in drug discovery projects. These shortcomings hinder their application in model building for drug discovery. To address this issue, we propose a multi-agent data mining system based on Large Language Models that effectively identifies experimental conditions within 14,401 bioassays. This approach facilitates merging entries from different sources, culminating in the creation of PharmaBench. Additionally, we have developed a data processing workflow to integrate data from various sources, resulting in 156,618 raw entries. Through this workflow, we constructed PharmaBench, a comprehensive benchmark set for ADMET properties, which comprises eleven ADMET datasets and 52,482 entries. This benchmark set is designed to serve as an open-source dataset for the development of AI models relevant to drug discovery projects.

Practical feature filter strategy to machine learning for small datasets in chemistry

Many potential use cases for machine learning in chemistry and materials science suffer from small dataset sizes, which demands special care for the model design in order to deliver reliable predictions. Hence, feature selection as the key determinant for dataset design is essential here. We propose a practical and efficient feature filter strategy to determine the best input feature candidates. We illustrate this strategy for the prediction of adsorption energies based on a public dataset and sublimation enthalpies using an in-house training dataset. The input of adsorption energies reduces the feature space from 12 dimensions to two and still delivers accurate results. For the sublimation enthalpies, three input configurations are filtered from 14 possible configurations with different dimensions for further productive predictions as being most relevant by using our feature filter strategy. The best extreme gradient boosting regression model possesses a good performance and is evaluated from statistical and theoretical perspectives, reaching a level of accuracy comparable to density functional theory computations and allowing for physical interpretations of the predictions. Overall, the results indicate that the feature filter strategy can help interdisciplinary scientists without rich professional AI knowledge and limited computational resources to establish a reliable small training dataset first, which may make the final machine learning model training easier and more accurate, avoiding time-consuming hyperparameter explorations and improper feature selection.

Statistical Metamodel of Liner Acoustic Impedance Based on Neural Network and Probabilistic Learning for Small Datasets

The main novelty of this paper consists of presenting a statistical artificial neural network (ANN)-based model for a robust prediction of the frequency-dependent aeroacoustic liner impedance using an aeroacoustic computational model (ACM) dataset of small size. The model, focusing on percentage of open area (POA) and sound pressure level (SPL) at a zero Mach number, takes into account uncertainties using a probabilistic formulation. The main difficulty in training an ANN-based model is the small size of the ACM dataset. The probabilistic learning carried out using the probabilistic learning on manifolds (PLoM) algorithm addresses this difficulty as it allows constructing a very large training dataset from learning the probabilistic model from a small dataset. A prior conditional probability model is presented for the PCA-based statistical reduced representation of the frequency-sampled vector of the log-resistance and reactance. It induces some statistical constraints that are not straightforwardly taken into account when training such an ANN-based model by classical optimizations methods under constraints. A second novelty of this paper consists of presenting an alternate solution that involves using conditional statistics estimated with learned realizations from PLoM. A numerical example is presented.

Decision Tree Regression vs. Gradient Boosting Regressor Models for the Prediction of Hygroscopic Properties of Borassus Fruit Fiber

This research focuses on the environmental-friendly production of Borassus fruit fibers (BNF), its characterization, and hygroscopic properties determination via Dynamic Vapor Sorption (DVS). The experimental results obtained from the hygroscopic behavior analysis were used to create a primary dataset to train and test Decision Tree Regression (DTR) and Gradient Boosting Regressor (GBR) models. The created primary dataset comprised 294 observations, from which 80% were used to train the models, and the remaining 20% were used for the testing of the two models. The models exhibited high accuracy, easy interpretability on the small-size dataset, and flexibility with regards to the nature of the relationship between the input and output variable. Both models successfully predicted the hygroscopic behavior with the Gradient Boosting Regressor outperforming Decision Tree Regression by indicating values of 0.012, 0.109, 0.059, and 0.999 for MSE, RMSE, MAE, and R2, respectively, during the desorption of the BNF, and values of 0.012, 0.109, 0.059, and 0.999 for MSE, RMSE, MAE, and R2, respectively, during the desorption of the BNF. This suggests that the Gradient Boosting Regressor illustrated the maximum accuracy. The outcomes can be utilized to provide an alternative for traditional methods, which can often be costly and time-consuming by improving the engineering properties of BNF. The models can be used in the construction sector to lower costs as they are able to pinpoint elements influencing the characteristics for specific applications to grasp its various properties through the prediction of its hygroscopic properties.

Detection dataset of electric bicycles for lift control

Electric bicycle datasets for lift control have played a crucial role in the development of electric bicycle lift entry detection algorithms. However, existing datasets often encounter issues, including a small dataset size, a high false detection rate, and the absence of a publicly available benchmark dataset. This paper addresses these challenges by constructing a new electric bicycle detection dataset for lift control based on the research demands of the electric bicycle detection task for lift control. XHNet_EB, a car scene image dataset covering various types of electric bicycles, was constructed based on images taken by a camera from above. Segmented annotation was employed, framing the front and rear of the electric bicycle independently. This approach effectively addressed the problems of many irrelevant features and a high false detection rate caused by the use of whole-electric-bicycle annotations in mainstream datasets. AP_50, AP_75, the mAP, and the number of false detections in the images were used for quantitative analysis of the dataset. The experimental results indicated that the object detection accuracy of the XHNet_EB dataset was excellent. Quantitative analysis and evaluation of the number of false detections in images were conducted using four mainstream lightweight detection model algorithms. The results demonstrated that segmented annotation reduced the false detection rate more effectively than entire electric bicycle annotation. This study identified drawbacks in existing datasets. The dataset proposed in this paper overcomes the shortcomings of existing commercial data and solves problems such as the high false detection rate caused by the inclusion of many irrelevant features caused by the “whole-electric-bicycle annotation” method, which could help with the development of electric bicycle detection applications in lifts.

Breaking the data barrier: a review of deep learning techniques for democratizing AI with small datasets

Justifiably, while big data is the primary interest of research and public discourse, it is essential to acknowledge that small data remains prevalent. The same technological and societal forces that generate big datasets also produce a more significant number of small datasets. Contrary to the notion that more data is inherently superior, real-world constraints such as budget limitations and increased analytical complexity present critical challenges. Quality versus quantity trade-offs necessitate strategic decision-making, where small data often leads to quicker, more accurate, and cost-effective insights. Concentrating AI research, particularly in deep learning (DL), on big datasets exacerbates AI inequality, as tech giants such as Meta, Amazon, Apple, Netflix and Google (MAANG) can easily lead AI research due to their access to vast datasets, creating a barrier for small and mid-sized enterprises that lack similar access. This article addresses this imbalance by exploring DL techniques optimized for small datasets, offering a comprehensive review of historic and state-of-the-art DL models developed specifically for small datasets. This study aims to highlight the feasibility and benefits of these approaches, promoting a more inclusive and equitable AI landscape. Through a PRISMA-based literature search, 175+ relevant articles are identified and subsequently analysed based on various attributes, such as publisher, country, utilization of small dataset technique, dataset size, and performance. This article also delves into current DL models and highlights open research problems, offering recommendations for future investigations. Additionally, the article highlights the importance of developing DL models that effectively utilize small datasets, particularly in domains where data acquisition is difficult and expensive.

ASD-GANNet: A Generative Adversarial Network-Inspired Deep Learning Approach for the Classification of Autism Brain Disorder.

The classification of a pre-processed fMRI dataset using functional connectivity (FC)-based features is considered a challenging task because of the set of high-dimensional FC features and the small dataset size. To tackle this specific set of FC high-dimensional features and a small-sized dataset, we propose here a conditional Generative Adversarial Network (cGAN)-based dataset augmenter to first train the cGAN on computed connectivity features of NYU dataset and use the trained cGAN to generate synthetic connectivity features per category. After obtaining a sufficient number of connectivity features per category, a Multi-Head attention mechanism is used as a head for the classification. We name our proposed approach "ASD-GANNet", which is end-to-end and does not require hand-crafted features, as the Multi-Head attention mechanism focuses on the features that are more relevant. Moreover, we compare our results with the six available state-of-the-art techniques from the literature. Our proposed approach results using the "NYU" site as a training set for generating a cGAN-based synthetic dataset are promising. We achieve an overall 10-fold cross-validation-based accuracy of 82%, sensitivity of 82%, and specificity of 81%, outperforming available state-of-the art approaches. A sitewise comparison of our proposed approach also outperforms the available state-of-the-art, as out of the 17 sites, our proposed approach has better results in the 10 sites.

Integrated Transfer Learning and Multitask Learning Strategies to Construct Graph Neural Network Models for Predicting Bioaccumulation Parameters of Chemicals.

Accurate prediction of parameters related to the environmental exposure of chemicals is crucial for the sound management of chemicals. However, the lack of large data sets for training models may result in poor prediction accuracy and robustness. Herein, integrated transfer learning (TL) and multitask learning (MTL) was proposed for constructing a graph neural network (GNN) model (abbreviated as TL-MTL-GNN model) using n-octanol/water partition coefficients as a source domain. The TL-MTL-GNN model was trained to predict three bioaccumulation parameters based on enlarged data sets that cover 2496 compounds with at least one bioaccumulation parameter. Results show that the TL-MTL-GNN model outperformed single-task GNN models with and without the TL, as well as conventional machine learning models trained with molecular descriptors or fingerprints. Applicability domains were characterized by a state-of-the-art structure-activity landscape-based (abbreviated as ADSAL) methodology. The TL-MTL-GNN model coupled with the optimal ADSAL was employed to predict bioaccumulation parameters for around 60,000 chemicals, with more than 13,000 compounds identified as bioaccumulative chemicals. The high predictive accuracy and robustness of the TL-MTL-GNN model demonstrate the feasibility of integrating the TL and MTL strategy in modeling small-sized data sets. The strategy holds significant potential for addressing small data challenges in modeling environmental chemicals.

TabDEG: Classifying differentially expressed genes from RNA-seq data based on feature extraction and deep learning framework.

Traditional differential expression genes (DEGs) identification models have limitations in small sample size datasets because they require meeting distribution assumptions, otherwise resulting high false positive/negative rates due to sample variation. In contrast, tabular data model based on deep learning (DL) frameworks do not need to consider the data distribution types and sample variation. However, applying DL to RNA-Seq data is still a challenge due to the lack of proper labeling and the small sample size compared to the number of genes. Data augmentation (DA) extracts data features using different methods and procedures, which can significantly increase complementary pseudo-values from limited data without significant additional cost. Based on this, we combine DA and DL framework-based tabular data model, propose a model TabDEG, to predict DEGs and their up-regulation/down-regulation directions from gene expression data obtained from the Cancer Genome Atlas database. Compared to five counterpart methods, TabDEG has high sensitivity and low misclassification rates. Experiment shows that TabDEG is robust and effective in enhancing data features to facilitate classification of high-dimensional small sample size datasets and validates that TabDEG-predicted DEGs are mapped to important gene ontology terms and pathways associated with cancer.

Pixel-level classification of pigmented skin cancer lesions using multispectral autofluorescence lifetime dermoscopy imaging.

There is no clinical tool available to primary care physicians or dermatologists that could provide objective identification of suspicious skin cancer lesions. Multispectral autofluorescence lifetime imaging (maFLIM) dermoscopy enables label-free biochemical and metabolic imaging of skin lesions. This study investigated the use of pixel-level maFLIM dermoscopy features for objective discrimination of malignant from visually similar benign pigmented skin lesions. Clinical maFLIM dermoscopy images were acquired from 60 pigmented skin lesions before undergoing a biopsy examination. Random forest and deep neural networks classification models were explored, as they do not require explicit feature selection. Feature pools with either spectral intensity or bi-exponential maFLIM features, and a combined feature pool, were independently evaluated with each classification model. A rigorous cross-validation strategy tailored for small-size datasets was adopted to estimate classification performance. Time-resolved bi-exponential autofluorescence features were found to be critical for accurate detection of malignant pigmented skin lesions. The deep neural network model produced the best lesion-level classification, with sensitivity and specificity of 76.84%±12.49% and 78.29%±5.50%, respectively, while the random forest classifier produced sensitivity and specificity of 74.73%±14.66% and 76.83%±9.58%, respectively. Results from this study indicate that machine-learning driven maFLIM dermoscopy has the potential to assist doctors with identifying patients in real need of biopsy examination, thus facilitating early detection while reducing the rate of unnecessary biopsies.