Deep Learning Research Articles

Deep evidential learning for radiotherapy dose prediction

Background:As we navigate towards integrating deep learning methods in the real clinic, a safety concern lies in whether and how the model can express its own uncertainty when making predictions. In this work, we present a novel application of an uncertainty-quantification framework called Deep Evidential Learning in the domain of radiotherapy dose prediction. Method:Using medical images of the Open Knowledge-Based Planning Challenge dataset, we found that this model can be effectively harnessed to yield uncertainty estimates that inherited correlations with prediction errors upon completion of network training. This was achieved only after reformulating the original loss function for a stable implementation. Results:We found that (i)epistemic uncertainty was highly correlated with prediction errors, with various association indices comparable or stronger than those for Monte-Carlo Dropout and Deep Ensemble methods, (ii)the median error varied with uncertainty threshold much more linearly for epistemic uncertainty in Deep Evidential Learning relative to these other two conventional frameworks, indicative of a more uniformly calibrated sensitivity to model errors, (iii)relative to epistemic uncertainty, aleatoric uncertainty demonstrated a more significant shift in its distribution in response to Gaussian noise added to CT intensity, compatible with its interpretation as reflecting data noise. Conclusion:Collectively, our results suggest that Deep Evidential Learning is a promising approach that can endow deep-learning models in radiotherapy dose prediction with statistical robustness. We have also demonstrated how this framework leads to uncertainty heatmaps that correlate strongly with model errors, and how it can be used to equip the predicted Dose-Volume-Histograms with confidence intervals.

Diagnostic performance of deep learning for infectious keratitis: a systematic review and meta-analysis

Infectious keratitis (IK) is the leading cause of corneal blindness globally. Deep learning (DL) is an emerging tool for medical diagnosis, though its value in IK is unclear. We aimed to assess the diagnostic accuracy of DL for IK and its comparative accuracy with ophthalmologists. In this systematic review and meta-analysis, we searched EMBASE, MEDLINE, and clinical registries for studies related to DL for IK published between 1974 and July 16, 2024. We performed meta-analyses using bivariate models to estimate summary sensitivities and specificities. This systematic review was registered with PROSPERO (CRD42022348596). Of 963 studies identified, 35 studies (136,401 corneal images from >56,011 patients) were included. Most studies had low risk of bias (68.6%) and low applicability concern (91.4%) in all domains of QUADAS-2, except the index test domain. Against the reference standard of expert consensus and/or microbiological results (seven external validation studies; 10,675 images), the summary estimates (95% CI) for sensitivity and specificity of DL for IK were 86.2% (71.6-93.9) and 96.3% (91.5-98.5). From 28 internal validation studies (16,059 images), summary estimates for sensitivity and specificity were 91.6% (86.8-94.8) and 90.7% (84.8-94.5). Based on seven studies (4007 images), DL and ophthalmologists had comparable summary sensitivity [89.2% (82.2-93.6) versus 82.2% (71.5-89.5); P=0.20] and specificity [(93.2% (85.5-97.0) versus 89.6% (78.8-95.2); P=0.45]. DL models may have good diagnostic accuracy for IK and comparable performance to ophthalmologists. These findings should be interpreted with caution due to the image-based analysis that did not account for potential correlation within individuals, relatively homogeneous population studies, lack of pre-specification of DL thresholds, and limited external validation. Future studies should improve their reporting, data diversity, external validation, transparency, and explainability to increase the reliability and generalisability of DL models for clinical deployment. NIH, Wellcome Trust, MRC, Fight for Sight, BHP, and ESCRS.

Deep learning method to automatically diagnose periodontal bone loss and periodontitis stage in dental panoramic radiograph.

Artificial intelligence (AI) could be used as an automatically diagnosis method for dental disease due to its accuracy and efficiency. This research proposed a novel convolutional neural network (CNN)-based deep learning (DL) ensemble model for tooth position detection, tooth outline segmentation, tooth tissue segmentation, periodontal bone loss and periodontitis stage prediction using dental panoramic radiographs. The dental panoramic radiographs of 320 patients during the period January 2020 to December 2023 were collected in our dataset. All images were de-identified without private information. In total, 8462 teeth were included. The algorithms that DL ensemble model adopted include YOLOv8, Mask R-CNN, and TransUNet. The prediction results of DL method were compared with diagnosis of periodontists. The periodontal bone loss degree deviation between the DL method and ground truth drawn by the three periodontists was 5.28%. The overall PCC value of the DL method and the periodontists' diagnoses was 0.832 (P <​ 0.001). ​The ICC value was 0.806 (P <​ 0.001). The total diagnostic accuracy of the DL method was 89.45%. The proposed DL ensemble model in this study shows high accuracy and efficiency in radiographic detection and a valuable adjunct to periodontal diagnosis. The method has strong potential to enhance clinical professional performance and build more efficient dental health services. The DL method not only could help dentists for rapid and accurate auxiliary diagnosis and prevent medical negligence, but also could be used as a useful learning resource for inexperienced dentists and dental students.

Non-destructive estimation for Kyoho grape shelf-life using Vis/NIR hyperspectral imaging and deep learning algorithm

Grape shelf-life estimation is a substantial challenge for the grape industry. The objective of this study is to investigate the potential of grape shelf-life estimation using HSI technique and a deep learning algorithm. The visible and near-infrared (400.68–1001.61 nm) hyperspectral reflectance images data of grape samples was acquired and preprocessed with different spectral preprocessing methods. Additionally, a stacked denoising autoencoder (SDAE)-based deep learning algorithm was developed to extract deep features from pixel-level hyperspectral data of grapes, and then these features were used as inputs to establish support vector machine (SVM) models for estimating grape shelf-life. Furthermore, SVM, one-dimensional convolutional neural network (1D CNN) and long short-term memory (LSTM) models were used as traditional machine learning and end to end models for comparison. The results demonstrated that the SDAE-SVM model achieved reasonable recognition accuracy of 100 % and 98.125 % for the shelf-life of grapes in the training and test sets, respectively. The overall results suggested that SDAE-based deep learning method can be used as a powerful tool to deal with large-scale hyperspectral data as well as this research confirms the feasibility of non-destructive estimation for grapes shelf-life by the combination of HSI technique and deep learning method, which would provide a valuable guidance for shelf-life estimation of other postharvest fruit.

Assessing forecast performance of daily reference evapotranspiration: A comparison of equations, machine and deep learning using weather forecasts

An accurate forecast of short-term reference evapotranspiration (ET0) is crucial for effective farm and irrigation scheduling management. The techniques of forecasting ET0 have progressed from empirical equations to artificial intelligence-based models. However, gaps have remained in comprehensively evaluating the ET0 forecast performance discrepancies between various methods based on weather forecasts in addition to historical meteorological data, as well as investigating the effect of the amount of input variables on ET0 forecast accuracy. Here, the study evaluated the forecast accuracy of five conventional equations calibrated and six machine learning (ML) and deep learning (DL) models in predicting daily ET0 for a lead time of 1–7 days in the North China Plain (NCP). Comparative evaluations indicate that the Analytical Penman-Monteith (APM) is the optimum empirical equation for forecasting ET0, but its accuracy decreases as lead times increase. Conversely, ML and DL models consistently outperform empirical equations, characterized by lower mean absolute error (MAE) and root mean square error (RMSE) as well as correlation coefficient and Nash-Sutcliffe efficiency coefficient approaching 1 across all lead times. Notably, the Bidirectional Long short-term memory (Bi-LSTM) model outperforms other models, maintaining robust and effective forecast performance even on day 7. Moreover, with the increasing input variables, combining weather forecasts and historical meteorological data, the forecast performance of all models is improved, with significant enhancement observed in the Bi-LSTM model. This work provides valuable insights for choosing appropriate models to forecast ET0, holding essential in regional managing farm and irrigation systems.

Advancements in Data Augmentation and Transfer Learning: A Comprehensive Survey to Address Data Scarcity Challenges

Abstract: Deep Learning (DL) models have demonstrated remarkable proficiency in image classification and recognition tasks, surpassing human capabilities. The observed enhancement in performance can be attributed to the utilization of extensive datasets. Nevertheless, DL models have huge data requirements. Widening the learning capability of such models from limited samples even today remains a challenge, given the intrinsic constraints of small datasets. The trifecta of challenges, encompassing limited labeled datasets, privacy, poor generalization performance, and the costliness of annotations, further compounds the difficulty in achieving robust model performance. Overcoming the challenge of expanding the learning capabilities of Deep Learning models with limited sample sizes remains a pressing concern even today. To address this critical issue, our study conducts a meticulous examination of established methodologies, such as Data Augmentation and Transfer Learning, which offer promising solutions to data scarcity dilemmas. Data Augmentation, a powerful technique, amplifies the size of small datasets through a diverse array of strategies. These encompass geometric transformations, kernel filter manipulations, neural style transfer amalgamation, random erasing, Generative Adversarial Networks, augmentations in feature space, and adversarial and meta- learning training paradigms. : Furthermore, Transfer Learning emerges as a crucial tool, leveraging pre-trained models to facilitate knowledge transfer between models or enabling the retraining of models on analogous datasets. Through our comprehensive investigation, we provide profound insights into how the synergistic application of these two techniques can significantly enhance the performance of classification tasks, effectively magnifying scarce datasets. This augmentation in data availability not only addresses the immediate challenges posed by limited datasets but also unlocks the full potential of working with Big Data in a new era of possibilities in DL applications.

An Adaptive Framework for Traffic Congestion Prediction using Deep Learning

Aim and background: Congestion on China's roads has worsened in recent years due to the country's rapid economic development, rising urban population, rising private car ownership, inequitable traffic flow distribution, and growing local congestion. As cities expand, traffic congestion has become an unavoidable nuisance that endangers the safety and progress of its residents. Improving the utilization rate of municipal transportation facilities and relieving traffic congestion depend on a thorough and accurate identification of the current state of road traffic and necessitate anticipating road congestion in the city. Methodology: In this research, we suggest using a deep spatial and temporal graph convolutional network (DSGCN) to forecast the current state of traffic congestion. To begin, we grid out the transportation system to create individual regions for analysis. In this work, we abstract the grid region centers as nodes, and we use an adjacency matrix to signify the dynamic correlations between the nodes. Results and Discussion: The spatial correlation between regions is then captured utilizing a Graph Convolutional-Neural-Network (GCNN), while the temporal correlation is captured using a two-layer long and short-term feature model (DSTM). Conclusion: Finally, testing on real PeMS datasets shows that the DSGCN has superior performance than other baseline models and provides more accurate traffic congestion prediction.

Artificial intelligence-enabled electrocardiogram for mortality and cardiovascular risk estimation: a model development and validation study

Artificial intelligence (AI)-enabled electrocardiography (ECG) can be used to predict risk of future disease and mortality but has not yet been adopted into clinical practice. Existing model predictions do not have actionability at an individual patient level, explainability, or biological plausibi. We sought to address these limitations of previous AI-ECG approaches by developing the AI-ECG risk estimator (AIRE) platform. The AIRE platform was developed in a secondary care dataset (Beth Israel Deaconess Medical Center [BIDMC]) of 1 163 401 ECGs from 189 539 patients with deep learning and a discrete-time survival model to create a patient-specific survival curve with a single ECG. Therefore, AIRE predicts not only risk of mortality, but also time-to-mortality. AIRE was validated in five diverse, transnational cohorts from the USA, Brazil, and the UK (UK Biobank [UKB]), including volunteers, primary care patients, and secondary care patients. AIRE accurately predicts risk of all-cause mortality (BIDMC C-index 0·775, 95% CI 0·773-0·776; C-indices on external validation datasets 0·638-0·773), future ventricular arrhythmia (BIDMC C-index 0·760, 95% CI 0·756-0·763; UKB C-index 0·719, 95% CI 0·635-0·803), future atherosclerotic cardiovascular disease (0·696, 0·694-0·698; 0·643, 0·624-0·662), and future heart failure (0·787, 0·785-0·789; 0·768, 0·733-0·802). Through phenome-wide and genome-wide association studies, we identified candidate biological pathways for the prediction of increased risk, including changes in cardiac structure and function, and genes associated with cardiac structure, biological ageing, and metabolic syndrome. AIRE is an actionable, explainable, and biologically plausible AI-ECG risk estimation platform that has the potential for use worldwide across a wide range of clinical contexts for short-term and long-term risk estimation. British Heart Foundation, National Institute for Health and Care Research, and Medical Research Council.

BOATMAP: Bayesian Optimization Active Targeting for Monomorphic Arrhythmia Pace-mapping

Recent advances in machine learning and deep learning have presented new opportunities for learning to localize the origin of ventricular activation from 12-lead electrocardiograms (ECGs), an important step in guiding ablation therapies for ventricular tachycardia. Passively learning from population data is faced with challenges due to significant variations among subjects, and building a patient-specific model raises the open question of where to select pace-mapping data for training. This work introduces BOATMAP, a novel active learning approach designed to provide clinicians with interpretable guidance that progressively assists in locating the origin of ventricular activation from 12-lead ECGs. BOATMAP inverts the input–output relationship in traditional machine learning solutions to this problem and learns the similarity between a target ECG and a paced ECG as a function of the pacing site coordinates. Using Gaussian processes (GP) as a surrogate model, BOATMAP iteratively refines the estimated similarity landscape while providing suggestions to clinicians regarding the next optimal pacing site. Furthermore, it can incorporate constraints to avoid suggesting pacing in non-viable regions such as the core of the myocardial scar. Tested in a realistic simulation environment in various heart geometries and tissue properties, BOATMAP demonstrated the ability to accurately localize the origin of activation, achieving an average localization accuracy of 3.9±3.6mm with only 8.0±4.0 pacing sites. BOATMAP offers real-time interpretable guidance for accurate localization and enhancing clinical decision-making.

A Thorough Review of Deep Learning in Autism Spectrum Disorder Detection: From Data to Diagnosis

Background: Autism Spectrum Disorder (ASD) is a multifaceted neurodevelopmental condition with significant heterogeneity in its clinical presentation. Timely and precise identification of ASD is crucial for effective intervention and assistance. Recent advances in deep learning techniques have shown promise in enhancing the accuracy of ASD detection. Objective: This comprehensive review aims to provide an overview of various deep learning methods employed in detecting ASD, utilizing diverse neuroimaging modalities. We analyze a range of studies that use resting-state functional Magnetic Resonance Imaging (rsfMRI), structural MRI (sMRI), task-based fMRI (tfMRI), and electroencephalography (EEG). This paper aims to assess the effectiveness of these techniques based on criteria such as accuracy, sensitivity, specificity, and computational efficiency. Methods: We systematically review studies investigating ASD detection using deep learning across different neuroimaging modalities. These studies utilize various preprocessing tools, atlases, feature extraction techniques, and classification algorithms. The performance metrics of interest include accuracy, sensitivity, specificity, precision, F1-score, recall, and area under the curve (AUC). Results: The review covers a wide range of studies, each with its own dataset and methodology. Notable findings include a study employing rsfMRI data from ABIDE that achieved an accuracy of 80% using LeNet. Another study using rsfMRI data from ABIDE-II achieved an impressive accuracy of 95.4% with the ASGCN deep learning model. Studies utilizing different modalities, such as EEG and sMRI, also reported high accuracies ranging from 74% to 95%. Conclusion: Deep learning-based approaches for ASD detection have demonstrated significant potential across multiple neuroimaging modalities. These methods offer a more objective and data-driven approach to diagnosis, potentially reducing the subjectivity associated with clinical evaluations. However, challenges remain, including the need for larger and more diverse datasets, model interpretability, and clinical validation. The field of deep learning in ASD diagnosis continues to evolve, holding promise for early and accurate identification of individuals with ASD, which is crucial for timely intervention and support.

Capsular attention Conv-LSTM network (CACN): A deep learning structure for crop yield estimation based on multispectral imagery

Precise prediction of agricultural production output is crucial for farmers, policymakers, and the Farming-related industry. This article introduces a novel methodology to crop yield forecasting using a capsular neural network equipped with Conv-LSTM and attention mechanism. Our model combines the strengths of 3DCNN, and Conv-LSTM, which can capture the temporal dependencies and 3D features of crop yield data, and attention mechanism, which Can prioritize the most significant characteristics for making predictions. We evaluated CACN on a sizable collection of data of soybean crop yield in the United States from 2003 to 2019 and evaluated against various cutting-edge deep learning models. The outcomes indicate that our suggested approach surpasses other models in performance in terms of RMSE, correlation coefficient, and prediction error map. Specifically, our model achieved approximately 14 % improvement in terms of RMSE, compared to the state-of-the-art model Deep-Yield. Our model also demonstrated the ability to extract more meaningful features and capture the complex relationships between crop yield data and meteorological variables. Overall, our proposed method shows great potential for accurate and efficient crop yield forecasting and can be applied to other crops and regions.

Leveraging the learning focal point algorithm for emotional intelligence

&lt;span&gt;One of the secrets of the success of the education process is taking into account the learner’s feelings. That is, the teacher must be characterized by high emotional intelligence (EI) to understand the student’s feelings in order to facilitate the indoctrination process for him. Within the framework of the project to create a robot teacher, we had to add this feature because of its importance. In this article, we create a computer application that classifies students' emotions based on deep learning and learning focal point (LFP) algorithm by analyzing facial expressions. That is, the robot will be able to know whether the student is happy, excited, or sad in order to deal with him appropriately.&lt;/span&gt;

Machine vision-based detection of surface defects in cylindrical battery cases

Automotive 21700 series lithium batteries are prone to surface defects during production and transportation, thus affecting their performance, so we propose a full-surface defect detection method for battery cases based on the synthesis of traditional image processing and deep learning to address this problem. First, the mechanism of surface defects on a battery case is analysed, and the types of surface defects are summarized. A suitable platform for image acquisition is designed for the severely reflective surface. Since there is no publicly available defect dataset for cylindrical battery cases, a defect dataset is established, and the dataset is augmented and expanded via the traditional method and the ACGAN model. For the defects on the top of the battery case, a traditional image processing algorithm is used to combine the roundness and the area of the area pixels to make a comprehensive judgement. For the defects on the bottom and side of the battery case, after comparing and analysing a variety of deep learning networks, the YOLOv7 model is selected to address the data characteristics of the large experimental inputs and the small targeted defects. To improve the accuracy of the model to meet the needs of real-time detection in industry, the CA attention mechanism, the DYHEAD dynamic detector head, and the slicing-assisted super inference (SAHI) method are used, with a final map accuracy reaching 98.1 %, which is 2.7 % higher than that of the initial model. Compared with mainstream target detection algorithms, the algorithm in this paper has good detection performance for cylindrical battery case defect detection and can be better applied to real-time detection in industry.

A flexible 2.5D medical image segmentation approach with in-slice and cross-slice attention

Deep learning has become the de facto method for medical image segmentation, with 3D segmentation models excelling in capturing complex 3D structures and 2D models offering high computational efficiency. However, segmenting 2.5D images, characterized by high in-plane resolution but lower through-plane resolution, presents significant challenges. While applying 2D models to individual slices of a 2.5D image is feasible, it fails to capture the spatial relationships between slices. On the other hand, 3D models face challenges such as resolution inconsistencies in 2.5D images, along with computational complexity and susceptibility to overfitting when trained with limited data. In this context, 2.5D models, which capture inter-slice correlations using only 2D neural networks, emerge as a promising solution due to their reduced computational demand and simplicity in implementation. In this paper, we introduce CSA-Net, a flexible 2.5D segmentation model capable of processing 2.5D images with an arbitrary number of slices. CSA-Net features an innovative Cross-Slice Attention (CSA) module that effectively captures 3D spatial information by learning long-range dependencies between the center slice (for segmentation) and its neighboring slices. Moreover, CSA-Net utilizes the self-attention mechanism to learn correlations among pixels within the center slice. We evaluated CSA-Net on three 2.5D segmentation tasks: (1) multi-class brain MR image segmentation, (2) binary prostate MR image segmentation, and (3) multi-class prostate MR image segmentation. CSA-Net outperformed leading 2D, 2.5D, and 3D segmentation methods across all three tasks, achieving average Dice coefficients and HD95 values of 0.897 and 1.40 mm for the brain dataset, 0.921 and 1.06 mm for the prostate dataset, and 0.659 and 2.70 mm for the ProstateX dataset, demonstrating its efficacy and superiority. Our code is publicly available at: https://github.com/mirthAI/CSA-Net.

Target Signal Communication Detection of Black Flying UAVs Based on Deep Learning Algorithm

Background: Unmanned aerial vehicles (UAVs) are being widely used in many fields, such as national economy, social development, national defense, and security. Currently, the number of registered UAVs in China is far less than that of flying UAVs-the frequent occurrence of unsafe incidents. Objective: The phenomenon of UAVs flying undeclared and unapproved has caused more serious troubles to social public order and people's production and life. Methods: In this paper, to assist the public security department in detecting the phenomenon of UAV black flying, our team conducts a series of research based on the deep learning YOLOv5 (You Only Look Once) algorithm. Results: Firstly, the Vision Transformer mechanism is integrated to enhance the robustness of the model. Secondly, depth-separable convolution is introduced to reduce parameter redundancy. Finally, the SimAM attention-free mechanism and CBAM attention-free mechanism are combined to enhance the attention of small target UAVs. Conclusion: Through the analysis of UAV targets in video surveillance, the rapid identification of black-flying UAVs can be realized, the monitoring and early warning ability of UAVs in a specific area can be improved, and the loss of life and property of people can be reduced or saved as much as possible.

A systematic review of deep learning-based denoising for low-dose computed tomography from a perceptual quality perspective.

Low-dose computed tomography (LDCT) scans are essential in reducing radiation exposure but often suffer from significant image noise that can impair diagnostic accuracy. While deep learning approaches have enhanced LDCT denoising capabilities, the predominant reliance on objective metrics like PSNR and SSIM has resulted in over-smoothed images that lack critical detail. This paper explores advanced deep learning methods tailored specifically to improve perceptual quality in LDCT images, focusing on generating diagnostic-quality images preferred in clinical practice. We review and compare current methodologies, including perceptual loss functions and generative adversarial networks, addressing the significant limitations of current benchmarks and the subjective nature of perceptual quality evaluation. Through a systematic analysis, this study underscores the urgent need for developing methods that balance both perceptual and diagnostic quality, proposing new directions for future research in the field.

Deep Learning in DNA- and RNA-sequence Analysis: Advances and New Challenges

Deep Learning in DNA- and RNA-sequence Analysis: Advances and New Challenges

Adversarial attacks in signature verification: a deep learning approach

Handwritten signature recognition in forensic science is crucial for identity and document authentication. While serving as a legal representation of a person’s agreement or consent to the contents of a document, handwritten signatures de termine the authenticity of a document, identify forgeries, pinpoint the suspects and support other pieces of evidence like ink or document analysis. This work focuses on developing and evaluating a handwritten signature verification sys tem using a convolutional neural network (CNN) and emphasising the model’s efficacy using hand-crafted adversarial attacks. Initially, handwritten signatures have been collected from sixteen volunteers, each contributing ten samples, fol lowed by image normalization and augmentation to boost synthetic data samples and overcome the data scarcity. The proposed model achieved a testing accu racy of 91.35% using an 80:20 train-test split. Additionally, using the five-fold cross-validation, the model achieved a robust validation accuracy of nearly 98%. Finally, the introduction of manually constructed adversarial assaults on the sig nature images undermines the model’s accuracy, bringing the accuracy down to nearly 80%. This highlights the need to consider adversarial resilience while designing deep learning models for classification tasks. Exposing the model to real look-alike fake samples is critical while testing its robustness and refining the model using trial and error methods.

Aircraft Closed-Loop Dynamic System Identification in the Entire Flight Envelope Range Based on Deep Learning

Aircraft Closed-Loop Dynamic System Identification in the Entire Flight Envelope Range Based on Deep Learning

Forensic bone age assessment of hand and wrist joint MRI images in Chinese han male adolescents based on deep convolutional neural networks.

In Chinese criminal law, the ages of 12, 14, 16, and 18 years old play a significant role in the determination of criminal responsibility. In this study, we developed an epiphyseal grading system based on magnetic resonance image (MRI) of the hand and wrist for the Chinese Han population and explored the feasibility of employing deep learning techniques for bone age assessment based on MRI of the hand and wrist. This study selected 282 Chinese Han Chinese males aged 6.0-21.0 years old. In the course of our study, we proposed a novel deep learning model for extracting and enhancing MRI hand and wrist bone features to enhance the prediction of target MRI hand and wrist bone age and achieve precise classification of the target MRI and regression of bone age. The evaluation metric for the classification model including precision, specificity, sensitivity, and accuracy, while the evaluation metrics chosen for the regression model are MAE. The epiphyseal grading was used as a supervised method, which effectively solved the problem of unbalanced sample distribution, and the two experts showed strong consistency in the epiphyseal plate grading process. In the classification results, the accuracy in distinguishing between adults and minors was 91.1%, and the lowest accuracy in the three minor classifications (12, 14, and 16 years of age) was 94.6%, 91.1% and 96.4%, respectively. The MAE of the regression results was 1.24 years. In conclusion, the deep learning model proposed enabled the age assessment of hand and wrist bones based on MRI.