Performance Of Deep Learning Models Research Articles

Spatial-temporal deep learning model based on Similarity Principle for dock shared bicycles ridership prediction

Demand prediction plays a critical role in traffic research. The key challenge of traffic demand prediction lies in modeling the complex spatial dependencies and temporal dynamics. However, there is no mature and widely accepted concept to support the solution of the above challenge. Essentially, a prediction model combined with similar objects in temporal and spatial dimensions could obtain better performance. This paper proposes a concept called the Similarity-based Principle (SP), which is applied to improve the prediction performance of deep learning models in complex traffic scenarios. For the temporal components, the long-term temporal dynamics in contemporaneous historical data for ridership are extracted by the Stacked Autoencoder (SAE) method. For the spatial components, the activity-based spatial geographic information (ABG-information) is used to capture the spatial correlation of the traffic network, which is reflected in the daily activities of humans. Specifically, the SP is applied to a Spatio-temporal Graph Convolutional Neural Network (STGCNN) model. In the case study, the Similarity-based Principle Spatio-temporal Graph Convolutional Neural Network (SP-STGCNN) model predicts demand for bicycle sharing in San Francisco. The results show that the SP effectively improves the model's performance. The prediction accuracy is enhanced by up to 10.34% compared with STGCNN. For spatial relationships, the model using the geographic information attribute performs better than that using the road information attribute and the distance attribute. It is proved that the construction of the Spatio-temporal model-based similarity principle can improve the performance.

Improving Generalizability of PET DL Algorithms: List-Mode Reconstructions Improve DOTATATE PET Hepatic Lesion Detection Performance.

Deep learning (DL) algorithms used for DOTATATE PET lesion detection typically require large, well-annotated training datasets. These are difficult to obtain due to low incidence of gastroenteropancreatic neuroendocrine tumors (GEP-NETs) and the high cost of manual annotation. Furthermore, networks trained and tested with data acquired from site specific PET/CT instrumentation, acquisition and processing protocols have reduced performance when tested with offsite data. This lack of generalizability requires even larger, more diverse training datasets. The objective of this study is to investigate the feasibility of improving DL algorithm performance by better matching the background noise in training datasets to higher noise, out-of-domain testing datasets. 68Ga-DOTATATE PET/CT datasets were obtained from two scanners: Scanner1, a state-of-the-art digital PET/CT (GE DMI PET/CT; n = 83 subjects), and Scanner2, an older-generation analog PET/CT (GE STE; n = 123 subjects). Set1, the data set from Scanner1, was reconstructed with standard clinical parameters (5 min; Q.Clear) and list-mode reconstructions (VPFXS 2, 3, 4, and 5-min). Set2, data from Scanner2 representing out-of-domain clinical scans, used standard iterative reconstruction (5 min; OSEM). A deep neural network was trained with each dataset: Network1 for Scanner1 and Network2 for Scanner2. DL performance (Network1) was tested with out-of-domain test data (Set2). To evaluate the effect of training sample size, we tested DL model performance using a fraction (25%, 50% and 75%) of Set1 for training. Scanner1, list-mode 2-min reconstructed data demonstrated the most similar noise level compared that of Set2, resulting in the best performance (F1 = 0.713). This was not significantly different compared to the highest performance, upper-bound limit using in-domain training for Network2 (F1 = 0.755; p-value = 0.103). Regarding sample size, the F1 score significantly increased from 25% training data (F1 = 0.478) to 100% training data (F1 = 0.713; p < 0.001). List-mode data from modern PET scanners can be reconstructed to better match the noise properties of older scanners. Using existing data and their associated annotations dramatically reduces the cost and effort in generating these datasets and significantly improves the performance of existing DL algorithms. List-mode reconstructions can provide an efficient, low-cost method to improve DL algorithm generalizability.

Approximating Intermediate Feature Maps of Self-Supervised Convolution Neural Network to Learn Hard Positive Representations in Chest Radiography.

Recent advances in contrastive learning have significantly improved the performance of deep learning models. In contrastive learning of medical images, dealing with positive representation is sometimes difficult because some strong augmentation techniques can disrupt contrastive learning owing to the subtle differences between other standardized CXRs compared to augmented positive pairs; therefore, additional efforts are required. In this study, we propose intermediate feature approximation (IFA) loss, which improves the performance of contrastive convolutional neural networks by focusing more on positive representations of CXRs without additional augmentations. The IFA loss encourages the feature maps of a query image and its positive pair to resemble each other by maximizing the cosine similarity between the intermediate feature outputs of the original data and the positive pairs. Therefore, we used the InfoNCE loss, which is commonly used loss to address negative representations, and the IFA loss, which addresses positive representations, together to improve the contrastive network. We evaluated the performance of the network using various downstream tasks, including classification, object detection, and a generative adversarial network (GAN) inversion task. The downstream task results demonstrated that IFA loss can improve the performance of effectively overcoming data imbalance and data scarcity; furthermore, it can serve as a perceptual loss encoder for GAN inversion. In addition, we have made our model publicly available to facilitate access and encourage further research and collaboration in the field.

Deep learning model performance for identifying pediatric acute respiratory distress syndrome on chest radiographs

PurposePediatric acute respiratory distress syndrome (PARDS) is underrecognized in the pediatric intensive care unit and the interpretation of chest radiographs is a key step in identification. We sought to test the performance of a machine learning model to detect PARDS in a cohort of children with respiratory failure.Materials and methodsA convolutional neural network (CNN) model previously developed to detect ARDS on adult chest radiographs was applied to a cohort of children age 7 days to 18 years, admitted to the PICU, and mechanically ventilated through a tracheostomy, endotracheal tube or full-face non-invasive positive pressure mask between May 2016 and January 2017. Two pediatric critical care physicians and a pediatric radiologist reviewed chest radiographs to evaluate if the chest radiographs were consistent with ARDS (bilateral airspace disease) and PARDS (any airspace disease) and the CNN model was tested against clinicians.ResultsA total of 328 chest radiographs were evaluated from 66 patients. Clinicians identified 84% (276/328) of the radiographs as potentially consistent with PARDS. Inter-rater reliability between individual clinicians and between the model and clinicians was similar (Cohen’s kappa 0.48 [95% CI 0.37–0.59] and 0.45 [95% CI 0.33–0.57], respectively). The model was better at identifying PARDS (AUC 0.882, F1 0.897) than ARDS (AUC 0.842, F1 0.742) and had equivalent or better performance to individual clinicians.ConclusionsAn ARDS detection model trained on adults performed well in detecting PARDS in children. Computer-assisted identification of PARDS on chest radiographs could improve the diagnosis of PARDS for enrollment in clinical trials and application of PARDS guidelines through improved diagnosis.

An early discovery of intrusion attack using novel optimized deep learning for internet of things

In the past couple of years, neural networks have gained widespread use in network security analysis. This type of analysis is usually performed in a nonlinear and highly correlated manner. Due to the immense amount of data traffic, the current models are prone to false alarms and poor detection. Deep-learning models can help security researchers identify and extract data features that are related to an attack. They can also minimize the data’s dimensionality and detect intrusions. Unfortunately, the complexity of the network structure and hidden neurons of a deep-learning model can be set by error-prone procedures. In order to improve the performance of deep learning models, a new algorithm is proposed. This method combines a gradient boost regression and particle swarm optimization. The proposes a method called the Spark-DBN-SVM-GBR algorithm. The simulations conducted proposed algorithm revealed that it has a better accuracy rate than other deep learning models and the experiments conducted on the PSO-GBR algorithm revealed that it performed better than the current optimization technique when detecting unauthorized attack activities.

Deep learning in oral cancer- a systematic review

BackgroundOral cancer is a life-threatening malignancy, which affects the survival rate and quality of life of patients. The aim of this systematic review was to review deep learning (DL) studies in the diagnosis and prognostic prediction of oral cancer.MethodsThis systematic review was conducted following the PRISMA guidelines. Databases (Medline via PubMed, Google Scholar, Scopus) were searched for relevant studies, from January 2000 to June 2023.ResultsFifty-four qualified for inclusion, including diagnostic (n = 51), and prognostic prediction (n = 3). Thirteen studies showed a low risk of biases in all domains, and 40 studies low risk for concerns regarding applicability. The performance of DL models was reported of the accuracy of 85.0–100%, F1-score of 79.31 - 89.0%, Dice coefficient index of 76.0 - 96.3% and Concordance index of 0.78–0.95 for classification, object detection, segmentation, and prognostic prediction, respectively. The pooled diagnostic odds ratios were 2549.08 (95% CI 410.77–4687.39) for classification studies.ConclusionsThe number of DL studies in oral cancer is increasing, with a diverse type of architectures. The reported accuracy showed promising DL performance in studies of oral cancer and appeared to have potential utility in improving informed clinical decision-making of oral cancer.

Establishment and comparison of in situ detection models for foodborne pathogen contamination on mutton based on SWIR-HSI.

Rapid and accurate detection of food-borne pathogens on mutton is of great significance to ensure the safety of mutton and its products and the health of consumers. The feasibility of short-wave infrared hyperspectral imaging (SWIR-HSI) in detecting the contamination status and species of Escherichia coli (EC), Staphylococcus aureus (SA) and Salmonella typhimurium (ST) contaminated on mutton was explored. The hyperspectral images of uncontaminated and contaminated mutton samples with different concentrations (108, 107, 106, 105, 104, 103 and 102 CFU/mL) of EC, SA and ST were acquired. The one dimensional convolutional neural network (1D-CNN) model was constructed and the influence of structure hyperparameters on the model was explored. The effects of different spectral preprocessing methods on partial least squares-discriminant analysis (PLS-DA), support vector machine (SVM) and 1D-CNN models were discussed. In addition, the feasibility of using the characteristic wavelength to establish simplified models was explored. The best full band model was the 1D-CNN model with the convolution kernels number of (64, 16) and the activation function of tanh established by the original spectra, and its accuracy of training set, test set and external validation set were 100.00, 92.86 and 97.62%, respectively. The optimal simplified model was genetic algorithm optimization support vector machine (GA-SVM). For discriminating the pathogen species, the accuracies of SVM models established by full band spectra preprocessed by 2D and all 1D-CNN models with the convolution kernel number of (32, 16) and the activation function of tanh were 100.00%. In addition, the accuracies of all simplified models were 100.00% except for the 1D-CNN models. Considering the complexity of features and model calculation, the 1D-CNN models established by original spectra were the optimal models for pathogenic bacteria contamination status and species. The simplified models provide basis for developing multispectral detection instruments. The results proved that SWIR-HSI combined with machine learning and deep learning could accurately detect the foodborne pathogen contamination on mutton, and the performance of deep learning models were better than that of machine learning. This study can promote the application of HSI technology in the detection of foodborne pathogens on meat.

Brain tumor classification from MRI scans: a framework of hybrid deep learning model with Bayesian optimization and quantum theory-based marine predator algorithm.

Brain tumor classification is one of the most difficult tasks for clinical diagnosis and treatment in medical image analysis. Any errors that occur throughout the brain tumor diagnosis process may result in a shorter human life span. Nevertheless, most currently used techniques ignore certain features that have particular significance and relevance to the classification problem in favor of extracting and choosing deep significance features. One important area of research is the deep learning-based categorization of brain tumors using brain magnetic resonance imaging (MRI). This paper proposes an automated deep learning model and an optimal information fusion framework for classifying brain tumor from MRI images. The dataset used in this work was imbalanced, a key challenge for training selected networks. This imbalance in the training dataset impacts the performance of deep learning models because it causes the classifier performance to become biased in favor of the majority class. We designed a sparse autoencoder network to generate new images that resolve the problem of imbalance. After that, two pretrained neural networks were modified and the hyperparameters were initialized using Bayesian optimization, which was later utilized for the training process. After that, deep features were extracted from the global average pooling layer. The extracted features contain few irrelevant information; therefore, we proposed an improved Quantum Theory-based Marine Predator Optimization algorithm (QTbMPA). The proposed QTbMPA selects both networks' best features and finally fuses using a serial-based approach. The fused feature set is passed to neural network classifiers for the final classification. The proposed framework tested on an augmented Figshare dataset and an improved accuracy of 99.80%, a sensitivity rate of 99.83%, a false negative rate of 17%, and a precision rate of 99.83% is obtained. Comparison and ablation study show the improvement in the accuracy of this work.

Deep Learning Model Effectiveness in Forecasting Limited-Size Aboveground Vegetation Biomass Time Series: Kenyan Grasslands Case Study

Timely forecasting of aboveground vegetation biomass is crucial for effective management and ensuring food security. However, research on predicting aboveground biomass remains scarce. Artificial intelligence (AI) methods could bridge this research gap and provide early warning to planners and stakeholders. This study evaluates the effectiveness of deep learning (DL) algorithms in predicting aboveground vegetation biomass with limited-size data. It employs an iterative forecasting procedure for four target horizons, comparing the performance of DL models—multi-layer perceptron (MLP), long short-term memory (LSTM), gated recurrent unit (GRU), convolutional neural network (CNN), and CNN-LSTM—against the traditional seasonal autoregressive integrated moving average (SARIMA) model, serving as a benchmark. Five limited-size vegetation biomass time series from Kenyan grasslands with values at 15-day intervals over a 20-year period were chosen for this purpose. Comparing the outcomes of these models revealed significant differences (p &lt; 0.05); however, none of the models proved superior among the five time series and the four horizons evaluated. The SARIMA, CNN, and CNN-LSTM models performed best, with the statistical model slightly outperforming the other two. Additionally, the accuracy of all five models varied significantly according to the prediction horizon (p &lt; 0.05). As expected, the accuracy of the models decreased as the prediction horizon increased, although this relationship was not strictly monotonic. Finally, this study indicated that, in limited-size aboveground vegetation biomass time series, there is no guarantee that deep learning methods will outperform traditional statistical methods.

Identifying Diabetic Retinopathy in the Human Eye: A Hybrid Approach Based on a Computer-Aided Diagnosis System Combined with Deep Learning.

Diagnosing and screening for diabetic retinopathy is a well-known issue in the biomedical field. A component of computer-aided diagnosis that has advanced significantly over the past few years as a result of the development and effectiveness of deep learning is the use of medical imagery from a patient's eye to identify the damage caused to blood vessels. Issues with unbalanced datasets, incorrect annotations, a lack of sample images, and improper performance evaluation measures have negatively impacted the performance of deep learning models. Using three benchmark datasets of diabetic retinopathy, we conducted a detailed comparison study comparing various state-of-the-art approaches to address the effect caused by class imbalance, with precision scores of 93%, 89%, 81%, 76%, and 96%, respectively, for normal, mild, moderate, severe, and DR phases. The analyses of the hybrid modeling, including CNN analysis and SHAP model derivation results, are compared at the end of the paper, and ideal hybrid modeling strategies for deep learning classification models for automated DR detection are identified.

Enhancing YOLO5 for the Assessment of Irregular Pelvic Radiographs with Multimodal Information.

Developmental dysplasia of the hip (DDH) is one of the most common orthopedic disorders in infants and young children. Accurate identification and localization of anatomical landmarks are prerequisites for the diagnosis of DDH. In recent years, various works have employed deep learning algorithms on radiography images for DDH diagnosis. However, none of these works have considered the incorporation of multimodal information. The pelvis exhibits distinct structures at different developmental stages, and there are also gender-based differences. In light of this, this study proposes a method to enhance the performance of deep learning models in diagnosing DDH by incorporating age and gender information into the channels. The study utilizes YOLO5 to construct a deep learning network for detecting hip joint landmarks. Moreover, a comprehensive dataset of 7750 pelvic X-ray images is established, covering ages from 4months to 16years and encompassing various conditions, such as deformities and post-operative cases, which authentically capture the temporal diversity and pathological complexities of DDH. Experimental results show that the YOLO5 model with integrated multimodal information achieves a mAP0.5-0.95 of 83.1% and a diagnostic accuracy of 86.7% in test dataset. The F1 scores for diagnosing cases of normal (NM), suspected dislocation (SD), mild dislocation (MD), and heavily dislocation (HD) are 90.9%, 79.8%, 63.5%, and 97.4%, respectively. Furthermore, experiments conducted on datasets of different sizes and networks of different sizes demonstrate the beneficial impact of multimodal information in improving the effectiveness of deep learning in diagnosing DDH.

Deep Learning Models Used in the Diagnostic Workup of Keratoconus: A Systematic Review and Exploratory Meta-Analysis.

The prevalence of keratoconus in the general population is reported to be up to 1 of 84. Over the past 2 decades, diagnosis and management evolved rapidly, but keratoconus screening in clinical practice is still challenging and asks for improving the accuracy of keratoconus detection. Deep learning (DL) offers considerable promise for improving the accuracy and speed of medical imaging interpretation. We establish an inventory of studies conducted with DL algorithms that have attempted to diagnose keratoconus. This systematic review was conducted according to the recommendations of the PRISMA statement. We searched (Pre-)MEDLINE, Embase, Science Citation Index, Conference Proceedings Citation Index, arXiv document server, and Google Scholar from inception to February 18, 2022. We included studies that evaluated the performance of DL algorithms in the diagnosis of keratoconus. The main outcome was diagnostic performance measured as sensitivity and specificity, and the methodological quality of the included studies was assessed using QUADAS-2. Searches retrieved 4100 nonduplicate records, and we included 19 studies in the qualitative synthesis and 10 studies in the exploratory meta-analysis. The overall study quality was limited because of poor reporting of patient selection and the use of inadequate reference standards. We found a pooled sensitivity of 97.5% (95% confidence interval, 93.6%-99.0%) and a pooled specificity of 97.2% (95% confidence interval, 85.7%-99.5%) for topography images as input. Our systematic review found that the overall diagnostic performance of DL models to detect keratoconus was good, but the methodological quality of included studies was modest.

Retracted: Enhance-Net: An Approach to Boost the Performance of Deep Learning Model Based on Real-Time Medical Images

Retracted: Enhance-Net: An Approach to Boost the Performance of Deep Learning Model Based on Real-Time Medical Images

Evaluation of Inference Performance of Deep Learning Models for Real-Time Weed Detection in an Embedded Computer.

The knowledge that precision weed control in agricultural fields can reduce waste and increase productivity has led to research into autonomous machines capable of detecting and removing weeds in real time. One of the driving factors for weed detection is to develop alternatives to herbicides, which are becoming less effective as weed species develop resistance. Advances in deep learning technology have significantly improved the robustness of weed detection tasks. However, deep learning algorithms often require extensive computational resources, typically found in powerful computers that are not suitable for deployment in robotic platforms. Most ground rovers and UAVs utilize embedded computers that are portable but limited in performance. This necessitates research into deep learning models that are computationally lightweight enough to function in embedded computers for real-time applications while still maintaining a base level of detection accuracy. This paper evaluates the weed detection performance of three real-time-capable deep learning models, YOLOv4, EfficientDet, and CenterNet, when run on a deep-learning-enabled embedded computer, an Nvidia Jetson Xavier AGX. We tested the accuracy of the models in detecting 13 different species of weeds and assesses their real-time viability through their inference speeds on an embedded computer compared to a powerful deep learning PC. The results showed that YOLOv4 performed better than the other models, achieving an average inference speed of 80 ms per image and 14 frames per second on a video when run on an imbedded computer, while maintaining a mean average precision of 93.4% at a 50% IoU threshold. Furthermore, recognizing that some real-world applications may require even greater speed, and that the detection program would not be the only task running on the embedded computer, a lightweight version of the YOLOv4 model, YOLOv4-tiny, was tested for improved performance in an embedded computer. YOLOv4-tiny impressively achieved an average inference speed of 24.5 ms per image and 52 frames per second, albeit with a slightly reduced mean average precision of 89% at a 50% IoU threshold, making it an ideal choice for real-time weed detection.

A deep learning-based illumination transform for devignetting photographs of dermatological lesions

Photographs of skin lesions taken with standard digital cameras (macroscopic images) have gained wide acceptance in dermatology. However, uneven background lighting caused by nonstandard image acquisition negatively impacts lesion segmentation and diagnosis. To address this, we propose an automated illumination equalization method based on a counter exponential transform (IECET). A modified residual network (ResNet) regressor is used to automate the selection of the operational parameter of the IECET. The regressor is designed by modifying the final fully-connected layer of the baseline ResNet-50 model. The modified fully-connected layer is coupled to a regression layer in the modified ResNet regressor. A prior knowledge base is created to train the modified ResNet regressor. For this, a set of corrupted images are generated by simulating uneven background illumination on pristine images. The knowledge base is created by including pairs of value components obtained from the HSV color space version of the corrupted macroscopic images and ideal operational parameter values that maximize the peak signal-to-noise ratio (PSNR) between the pristine images and the IECET outputs. We evaluated segmentation accuracies of the deep threshold prediction network (DTP-Net), DeepLabV3+, fully convolutional network (FCN), and U-Net on the corrupted macroscopic images and output images of the IECET. The DTP-Net, DeepLabV3+, FCN, and U-Net exhibited Dice similarity coefficient (DSC) of 0.71 ± 0.26, 0.85 ± 0.15, 0.75 ± 0.22, and 0.66 ± 0.28 on corrupted images and 0.81 ± 0.17, 0.87 ± 0.12, 0.79 ± 0.18, and 0.79 ± 0.15, on the outputs of the IECET. Increase in DSC proves the ability of the IECET to improve the performance of deep learning models used to segment skin lesions on macroscopic images.

Deep learning-based identification of eyes at risk for glaucoma surgery

To develop and evaluate the performance of a deep learning model (DLM) that predicts eyes at high risk of surgical intervention for uncontrolled glaucoma based on multimodal data from an initial ophthalmology visit. Longitudinal, observational, retrospective study. 4898 unique eyes from 4038 adult glaucoma or glaucoma-suspect patients who underwent surgery for uncontrolled glaucoma (trabeculectomy, tube shunt, xen, or diode surgery) between 2013 and 2021, or did not undergo glaucoma surgery but had 3 or more ophthalmology visits. We constructed a DLM to predict the occurrence of glaucoma surgery within various time horizons from a baseline visit. Model inputs included spatially oriented visual field (VF) and optical coherence tomography (OCT) data as well as clinical and demographic features. Separate DLMs with the same architecture were trained to predict the occurrence of surgery within 3 months, within 3–6 months, within 6 months–1 year, within 1–2 years, within 2–3 years, within 3–4 years, and within 4–5 years from the baseline visit. Included eyes were randomly split into 60%, 20%, and 20% for training, validation, and testing. DLM performance was measured using area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC). Shapley additive explanations (SHAP) were utilized to assess the importance of different features. Model prediction of surgery for uncontrolled glaucoma within 3 months had the best AUC of 0.92 (95% CI 0.88, 0.96). DLMs achieved clinically useful AUC values (> 0.8) for all models that predicted the occurrence of surgery within 3 years. According to SHAP analysis, all 7 models placed intraocular pressure (IOP) within the five most important features in predicting the occurrence of glaucoma surgery. Mean deviation (MD) and average retinal nerve fiber layer (RNFL) thickness were listed among the top 5 most important features by 6 of the 7 models. DLMs can successfully identify eyes requiring surgery for uncontrolled glaucoma within specific time horizons. Predictive performance decreases as the time horizon for forecasting surgery increases. Implementing prediction models in a clinical setting may help identify patients that should be referred to a glaucoma specialist for surgical evaluation.

Enhancing The Accuracy of Image Classification Using Deep Learning and Preprocessing Methods

Deep learning is one of many methods in Artificial Intelligence (AI) that computers can use to process information like text, images, and audio. This manuscript will be focusing on image preprocessing, one of the many different techniques that are used to modify the neural network model training process, and how it affects the training speed and accuracy of the neural network. Six different image preprocessing techniques were picked for use in this study: Grayscale, Smoothing, Unmask Sharpening, Laplacian and Equalization, and Random Cropping and Rotation all of which were implemented using Python and the libraries NumPy, OpenCV, and PyTorch. For the dataset, a batch of 10000 images from the CIFAR10 dataset were used to train the model. This study explored the impact of preprocessing techniques on a deep learning model, employing the RESNET50 architecture. Notable improvements in model accuracy were observed, particularly with normalization and random cropping accompanied by rotation. The efficiency gains attributed to preprocessing were highlighted, leading to a more rapid training process and significant resource savings. This research underscores the importance of thoughtful preprocessing in enhancing the performance of deep learning models, offering valuable insights for practitioners in imageclassification tasks.

Composition of Feature Selection for Time-Series Prediction with Deep Learning

The concept of predictions has gained much attention over the last few years. Research on prediction based on experience is error-prone. Usually, a lot of data has been available with multiple variables, but not all features are relevant. This work uses different feature selection techniques and their combinations to select the best possible feature subset. This paper uses feature selection techniques for time series data prediction with deep learning models. In this work, we have used filter, wrapper, and embedded methods for feature selection. We combine these techniques and compare their performance with single feature selection techniques. We found that the embedded method alone performs well with fewer features, but its performance deteriorates when the number of features increases. We input these reduced feature sets into deep learning models such as recurrent neural networks (RNN), long short-term memory (LSTM), and 1d Convolutional with LSTM (1d CNN-LSTM). We assess the performance of different deep-learning models with reduced features. Our results analysis found that the right feature subset may help increase the performance of deep learning models for time series prediction. Performance improvement is specific to the dataset, feature selection techniques, and the deployed DL models; performance improvement varies on average, from a few percent to fifty percent.

The impact of data augmentation and transfer learning on the performance of deep learning models for the segmentation of the hip on 3D magnetic resonance images

Different pathologies of the hip are characterized by the abnormal shape of the bony structures of the joint, namely the femur and the acetabulum. Three-dimensional (3D) models of the hip can be used for diagnosis, biomechanical simulation, and planning of surgical treatments. These models can be generated by building 3D surfaces of the joint's structures segmented on magnetic resonance (MR) images. Deep learning can avoid time-consuming manual segmentations, but its performance depends on the amount and quality of the available training data. Data augmentation and transfer learning are two approaches used when there is only a limited number of datasets. In particular, data augmentation can be used to artificially increase the size and diversity of the training datasets, whereas transfer learning can be used to build the desired model on top of a model previously trained with similar data. This study investigates the effect of data augmentation and transfer learning on the performance of deep learning for the automatic segmentation of the femur and acetabulum on 3D MR images of patients diagnosed with femoroacetabular impingement. Transfer learning was applied starting from a model trained for the segmentation of the bony structures of the shoulder joint, which bears some resemblance to the hip joint. Our results suggest that data augmentation is more effective than transfer learning, yielding a Dice similarity coefficient compared to ground-truth manual segmentations of 0.84 and 0.89 for the acetabulum and femur, respectively, whereas the Dice coefficient was 0.78 and 0.88 for the model based on transfer learning. The Accuracy for the two anatomical regions was 0.95 and 0.97 when using data augmentation, and 0.87 and 0.96 when using transfer learning. Data augmentation can improve the performance of deep learning models by increasing the diversity of the training dataset and making the models more robust to noise and variations in image quality. The proposed segmentation model could be combined with radiomic analysis for the automatic evaluation of hip pathologies.

Spectraformer: deep learning model for grain spectral qualitative analysis based on transformer structure.

This study delves into the use of compact near-infrared spectroscopy instruments for distinguishing between different varieties of barley, chickpeas, and sorghum, addressing a vital need in agriculture for precise crop variety identification. This identification is crucial for optimizing crop performance in diverse environmental conditions and enhancing food security and agricultural productivity. We also explore the potential application of transformer models in near-infrared spectroscopy and conduct an in-depth evaluation of the impact of data preprocessing and machine learning algorithms on variety classification. In our proposed spectraformer multi-classification model, we successfully differentiated 24 barley varieties, 19 chickpea varieties, and ten sorghum varieties, with the highest accuracy rates reaching 85%, 95%, and 86%, respectively. These results demonstrate that small near-infrared spectroscopy instruments are cost-effective and efficient tools with the potential to advance research in various identification methods, but also underscore the value of transformer models in near-infrared spectroscopy classification. Furthermore, we delve into the discussion of the influence of data preprocessing on the performance of deep learning models compared to traditional machine learning models, providing valuable insights for future research in this field.