Deep Learning System Research Articles

Performance of a Deep Learning System and Performance of Optometrists for the Detection of Glaucomatous Optic Neuropathy Using Colour Retinal Photographs

Background/Objectives: Glaucoma is the leading cause of irreversible blindness, with a significant proportion of cases remaining undiagnosed globally. The interpretation of optic disc and retinal nerve fibre layer images poses challenges for optometrists and ophthalmologists, often leading to misdiagnosis. AI has the potential to improve diagnosis. This study aims to validate an AI system (a convolutional neural network based on the Inception-v3 architecture) for detecting glaucomatous optic neuropathy (GON) using colour fundus photographs from a UK population and to compare its performance against Australian optometrists. Methods: A retrospective external validation study was conducted, comparing AI’s performance with that of 11 AHPRA-registered optometrists in Australia on colour retinal photographs, evaluated against a reference (gold) standard established by a panel of glaucoma specialists. Statistical analyses were performed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). Results: For referable GON, the sensitivity of the AI (33.3% [95%CI: 32.4–34.3) was significantly lower than that of optometrists (65.1% [95%CI: 64.1–66.0]), p < 0.0001, although with significantly higher specificity (AI: 97.4% [95%CI: 97.0–97.7]; optometrists: 85.5% [95%CI: 84.8–86.2], p < 0.0001). The optometrists demonstrated significantly higher AUROC (0.753 [95%CI: 0.744–0.762]) compared to AI (0.654 [95%CI: 0.645–0.662], p < 0.0001). Conclusion: The AI system exhibited lower performance than optometrists in detecting referable glaucoma. Our findings suggest that while AI can serve as a screening tool, both AI and optometrists have suboptimal performance for the nuanced diagnosis of glaucoma using fundus photographs alone. Enhanced training with diverse populations for AI is essential for improving GON detection and addressing the significant challenge of undiagnosed cases.

Effective automatic classification methods via deep learning for myopic maculopathy

BackgroundPathologic myopia (PM) associated with myopic maculopathy (MM) is a significant cause of visual impairment, especially in East Asia, where its prevalence has surged. Early detection and accurate classification of myopia-related fundus lesions are critical for managing PM. Traditional clinical analysis of fundus images is time-consuming and dependent on specialist expertise, driving the need for automated, accurate diagnostic tools.MethodsThis study developed a deep learning-based system for classifying five types of MM using color fundus photographs. Five architectures—ResNet50, EfficientNet-B0, Vision Transformer (ViT), Contrastive Language-Image Pre-Training (CLIP), and RETFound—were utilized. An ensemble learning approach with weighted voting was employed to enhance model performance. The models were trained on a dataset of 2,159 annotated images from Shenzhen Eye Hospital, with performance evaluated using accuracy, sensitivity, specificity, F1-Score, Cohen’s Kappa, and area under the receiver operating characteristic curve (AUC).ResultsThe ensemble model achieved superior performance across all metrics, with an accuracy of 95.4% (95% CI: 93.0–97.0%), sensitivity of 95.4% (95% CI: 86.8–97.5%), specificity of 98.9% (95% CI: 97.1–99.5%), F1-Score of 95.3% (95% CI: 93.2–97.2%), Kappa value of 0.976 (95% CI: 0.957–0.989), and AUC of 0.995 (95% CI: 0.992–0.998). The voting ensemble method demonstrated robustness and high generalization ability in classifying complex lesions, outperforming individual models.ConclusionThe ensemble deep learning system significantly enhances the accuracy and reliability of MM classification. This system holds potential for assisting ophthalmologists in early detection and precise diagnosis, thereby improving patient outcomes. Future work could focus on expanding the dataset, incorporating image quality assessment, and optimizing the ensemble algorithm for better efficiency and broader applicability.

A deep learning-based method for assessing tricuspid regurgitation using continuous wave Doppler spectra

Transthoracic echocardiography (TTE) is widely recognized as one of the principal modalities for diagnosing tricuspid regurgitation (TR). The diagnostic procedures associated with conventional methods are intricate and labor-intensive, with human errors leading to measurement variability, with outcomes critically dependent on the operators’ diagnostic expertise. In this study, we present an innovative assessment methodology for evaluating TR severity utilizing an end-to-end deep learning system. This deep learning system comprises a segmentation model of single cardiac cycle TR continuous wave (CW) Doppler spectra and a classification model of the spectra, trained on the TR CW Doppler spectra from a cohort of 11,654 patients. The efficacy of this intelligent assessment methodology was validated on 1500 internal cases and 573 external cases. The receiver operating characteristic (ROC) curves of the internal validation results indicate that the deep learning system achieved the areas under curve (AUCs) of 0.88, 0.84, and 0.89 for mild, moderate, and severe TR, respectively. The ROC curves of the external validation results demonstrate that the system attained the AUCs of 0.86, 0.79, and 0.87 for mild, moderate, and severe TR, respectively. Our study results confirm the feasibility and efficacy of this novel intelligent assessment method for TR severity.

Transforming Skin Cancer Diagnosis: A Deep Learning Approach with the Ham10000 Dataset

Skin cancer (SC) is one of the three most common cancers worldwide. Melanoma has the deadliest potential to spread to other parts of the body among all SCs. For SC treatments to be effective, early detection is essential. The high degree of similarity between tumor and non-tumors makes SC diagnosis difficult even for experienced doctors. To address this issue, authors have developed a novel Deep Learning (DL) system capable of automatically classifying skin lesions into seven groups: actinic keratosis (AKIEC), melanoma (MEL), benign keratosis (BKL), melanocytic Nevi (NV), basal cell carcinoma (BCC), dermatofibroma (DF), and vascular (VASC) skin lesions. Authors introduced the Multi-Grained Enhanced Deep Cascaded Forest (Mg-EDCF) as a novel DL model. In this model, first, researchers utilized subsampled multigrained scanning (Mg-sc) to acquire micro features. Second, authors employed two types of Random Forest (RF) to create input features. Finally, the Enhanced Deep Cascaded Forest (EDCF) was utilized for classification. The HAM10000 dataset was used for implementing, training, and evaluating the proposed and Transfer Learning (TL) models such as ResNet, AlexNet, and VGG16. During the validation and training stages, the performance of the four networks was evaluated by comparing their accuracy and loss. The proposed method outperformed the competing models with an average accuracy score of 98.19%. Our proposed methodology was validated against existing state-of-the-art algorithms from recent publications, resulting in consistently greater accuracies than those of the classifiers.

Automated acute skin toxicity scoring in a mouse model through deep learning.

This study presents a novel approach to skin toxicity assessment in preclinical radiotherapy trials through an advanced imaging setup and deep learning. Skin reactions, commonly associated with undesirable side effects in radiotherapy, were meticulously evaluated in 160 mice across four studies. A comprehensive dataset containing 7542 images was derived from proton/electron trials with matched manual scoring of the acute toxicity on the right hind leg, which was the target area irradiated in the trials. This dataset was the foundation for the subsequent model training. The two-step deep learning framework incorporated an object detection model for hind leg detection and a classification model for toxicity classification. An observer study involving five experts and the deep learning model, was conducted to analyze the retrospective capabilities and inter-observer variations. The results revealed that the hind leg object detection model exhibited a robust performance, achieving an accuracy of almost 99%. Subsequently, the classification model demonstrated an overall accuracy of about 85%, revealing nuanced challenges in specific toxicity grades. The observer study highlighted high inter-observer agreement and showcased the model's superiority in accuracy and misclassification distance. In conclusion, this study signifies an advancement in objective and reproducible skin toxicity assessment. The imaging and deep learning system not only allows for retrospective toxicity scoring, but also presents a potential for minimizing inter-observer variation and evaluation times, addressing critical gaps in manual scoring methodologies. Future recommendations include refining the system through an expanded training dataset, paving the way for its deployment in preclinical research and radiotherapy trials.

A Deep Learning System for Early Diabetic Retinopathy Diagnosis

A Deep Learning System for Early Diabetic Retinopathy Diagnosis

Motion controller for multi-joint robotic arm with deep cascade gated Bayesian broad learning system

Intelligent controllers based on the broad learning system can simplify the process of model parameter adjustment, finding wide applications in the motion control of multi-joint robotic arms. However, motion controllers for multi-joint robotic arms based on broad learning system exhibit insufficient precision and overlook the impact of joint motion commonalities on controller design. Therefore, this paper proposes a novel motion control strategy for a multi-joint robotic arm based on a deep cascade feature-enhancement gated Bayesian broad learning system. Firstly, the motion controller of the deep cascade feature-enhancement Bayesian broad learning system is constructed to enhance the robotic arm motion control precision. Secondly, an incremental node generation module with an attention-gated mechanism is constructed to capture the unique motion characteristics of the target joints, which is further combined with model generalization to simplify the motion control process of the multi-joint robotic arm. Finally, controller convergence is enhanced by combining it with the Lyapunov theory to constrain the learning parameters. Simulations and physical experiments are designed to verify the feasibility and superiority of the proposed motion control strategy. The results demonstrated that the strategy improved the accuracy of robotic arm motion control, with the root mean square error in position tracking reduced to 0.0019 rad. This represents a 93.39% reduction in error compared to existing techniques.

Developing Anticancer Drug Response System Using Deep Learning System with Hybrid Genomic and Chemical Features

Developing Anticancer Drug Response System Using Deep Learning System with Hybrid Genomic and Chemical Features

HyMNet: A Multimodal Deep Learning System for Hypertension Prediction Using Fundus Images and Cardiometabolic Risk Factors

Study Objectives: This study aimed to develop a multimodal deep learning (MMDL) system called HyMNet, integrating fundus images and cardiometabolic factors (age and sex) to enhance hypertension (HTN) detection. Methods: HyMNet employed RETFound, a model pretrained on 1.6 million retinal images, for the fundus data, in conjunction with a fully connected neural network for age and sex. The two pathways were jointly trained by joining their feature vectors into a fusion network. The system was trained on 5016 retinal images from 1243 individuals provided by the Saudi Ministry of National Guard Health Affairs. The influence of diabetes on HTN detection was also assessed. Results: HyMNet surpassed the unimodal system, achieving an F1 score of 0.771 compared to 0.745 for the unimodal model. For diabetic patients, the F1 score was 0.796, while it was 0.466 for non-diabetic patients. Conclusions: HyMNet exhibited superior performance relative to unimodal approaches, with an F1 score of 0.771 for HyMNet compared to 0.752 for models trained on demographic data alone, underscoring the advantages of MMDL systems in HTN detection. The findings indicate that diabetes significantly impacts HTN prediction, enhancing detection accuracy among diabetic patients. Utilizing MMDL with diverse data sources could improve clinical applicability and generalization.

Oral screening of dental calculus, gingivitis and dental caries through segmentation on intraoral photographic images using deep learning

ObjectiveIntraoral photographic images are instrumental in the early screening and clinical diagnosis of oral diseases. In addition, people have been trying to apply artificial intelligence to these images. The purpose of this study is to investigate and evaluate a deep learning system designed to segment intraoral photographic images for the detection of dental caries, dental calculus, and gingivitis, and to assess the degree of dental calculus based on the overall features of the tooth surface and gingival margin.Material and methodsThis cross-sectional study collected 3,365 oral endoscopic images, randomly distributed in training datasets (2,019 images), validation dataset (673 images), and test dataset (673 images). The training set and verification set images are manually labeled. An oral endoscopic image segmentation method based on Mamba (Oral-Mamba) and an intelligent evaluation model of dental calculus degree were proposed, achieving the segmentation of two types of oral diseases, namely gingivitis and dental caries, as well as the segmentation of dental calculus regions, and the intelligent evaluation of the degree of dental calculus.ResultsOral-Mamba demonstrated high accuracy in segmentation, with accuracy rates for gingivitis, dental caries, and dental calculus at 0.83, 0.83, and 0.81, respectively. In particular, these rates surpassed those of the U-Net model in IoU, accuracy, and recall metrics. Furthermore, Oral-Mamba runs 25% faster than U-Net.The accuracy of degree classification in the intelligent evaluation model of dental calculus degree is 85%.ConclusionThe proposed deep learning system is expected to be used for the detection of two types of oral diseases and dental calculus, and the degree judgment of photographic images from an intraoral camera. This system offers a practical method to assist in the oral screening of dental caries, dental calculus, and gingivitis, providing benefits such as intuitive use, time efficiency, cost-effectiveness, and ease of deployment.

Prediction of a Cephalometric Parameter and Skeletal Patterns from Lateral Profile Photographs: A Retrospective Comparative Analysis of Regression Convolutional Neural Networks.

Background/Objectives: Cephalometric analysis has a pivotal role in the quantification of the craniofacial skeletal complex, facilitating the diagnosis and management of dental malocclusions and underlying skeletal discrepancies. This study aimed to develop a deep learning system that predicts a cephalometric skeletal parameter directly from lateral profile photographs, potentially serving as a preliminary resource to motivate patients towards orthodontic treatment. Methods: ANB angle values and corresponding lateral profile photographs were obtained from the medical records of 1600 subjects (1039 female and 561 male, age range 3 years 8 months to 69 years 1 month). The lateral profile photographs were randomly divided into a training dataset (1250 images) and a test dataset (350 images). Seven regression convolutional neural network (CNN) models were trained on the lateral profile photographs and measured ANB angles. The performance of the models was assessed using the coefficient of determination (R2) and mean absolute error (MAE). Results: The R2 values of the seven CNN models ranged from 0.69 to 0.73, and the MAE values ranged from 1.46 to 1.53. Among the seven models, InceptionResNetV2 showed the highest success rate for predictions of ANB angle within 1° of range and the highest performance in skeletal class prediction, with macro-averaged accuracy, precision, recall, and F1 scores of 73.1%, 78.5%, 71.1%, and 73.0%, respectively. Conclusions: The proposed deep CNN models demonstrated the ability to predict a cephalometric skeletal parameter directly from lateral profile photographs, with 71% of predictions being within 2° of accuracy. This level of accuracy suggests potential clinical utility, particularly as a non-invasive preliminary screening tool. The system's ability to provide reasonably accurate predictions without radiation exposure could be especially beneficial for initial patient assessments and may enhance efficiency in orthodontic workflows.

A Pilot Study of Deep Learning-Based Monocular Depth Estimation from Fundus Photographs

The purpose of this study was to evaluate the feasibility of a generalizable deep-learning (DL) based system with no a priori knowledge of fundus photographs to generate monocular depth map information about optic disc structures from this imaging modality. Images of 30 stereo pairs of fundus photographs centered on the optic disc of 30 subjects were analyzed with this DL system to generate monocular depth maps using zero-shot cross-dataset transfer. These maps were registered onto reference standard depth maps derived from Optical Coherence Tomography. Accuracy of the DL system was assessed by the root of mean squared error (RMSE) between the estimate and reference standard. 47% of the total images from the dataset were successfully processed, with mean RMSE of 0.081. Our findings demonstrate that single image, monocular depth estimation with a generalizable DL system using zero-shot cross-dataset transfer applied to retinal color fundus photographs is feasible and has potential.

Real-time deep learning-assisted mechano-acoustic system for respiratory diagnosis and multifunctional classification

Epidermally mounted sensors using triaxial accelerometers have been previously used to monitor physiological processes with the implementation of machine learning (ML) algorithm interfaces. The findings from these previous studies have established a strong foundation for the analysis of high-resolution, intricate signals, typically through frequency domain conversion. In this study we integrate a wireless mechano-acoustic sensor with a multi-modal deep learning system for the real-time analysis of signals emitted by the laryngeal prominence area of the thyroid cartilage at frequency ranges up to 1 kHz. This interface provides real-time data visualization and communication with the ML server, creating a system that assesses severity of chronic obstructive pulmonary disease and analyzes the user’s speech patterns.

Deep Learning to Discriminate Arteritic From Nonarteritic Ischemic Optic Neuropathy on Color Images

Prompt and accurate diagnosis of arteritic anterior ischemic optic neuropathy (AAION) from giant cell arteritis and other systemic vasculitis can contribute to preventing irreversible vision loss from these conditions. Its clinical distinction from nonarteritic anterior ischemic optic neuropathy (NAION) can be challenging, especially when systemic symptoms are lacking or laboratory markers of the disease are not reliable. To develop, train, and test a deep learning system (DLS) to discriminate AAION from NAION on color fundus images during the acute phase. This was an international study including color fundus images of 961 eyes of 802 patients with confirmed AAION and NAION. Training was performed using images from 21 expert neuro-ophthalmology centers in 16 countries, while external testing was performed in a cohort from 5 expert neuro-ophthalmology centers in the US and Europe. Data for training and external testing were collected from August 2018 to January 2023. A mix of deidentified images of 2 fields of view (optic disc centered and macula centered) were used. For training and internal validation, images were from 16 fundus camera models with fields of 30° to 55°. For external testing, images were from 5 fundus cameras with fields of 30° to 50°. Data were analyzed from January 2023 to January 2024. The performance of the DLS was measured using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. In the training and validation sets, 374 (54.9%) of patients were female, 301 (44.2%) were male, and 6 (0.9%) were of unknown sex; the median (range) age was 66 (23-96) years. When tested on the external dataset including 121 patients (35 [28.9%] female, 44 [36.4%] male, and 42 [34.7%] of unknown sex; median [range] age, 69 [37-89] years), the DLS achieved an AUC of 0.97 (95% CI, 0.95-0.99), a sensitivity of 91.1% (95% CI, 85.2-96.9), a specificity of 93.4% (95% CI, 91.1-98.2), and an accuracy of 92.6% (95% CI, 90.5-96.6). The accuracy of the 2 experts for classification of the same dataset was 74.3% (95% CI, 66.7-81.9) and 81.6% (95% CI, 74.8-88.4), respectively. A DLS showing disease-specific averaged class-activation maps had greater than 90% accuracy at discriminating between acute AAION from NAION on color fundus images, at the eye level, without any clinical or biomarker information. A DLS that identifies AAION could improve clinical decision-making, potentially reducing the risk of misdiagnosis and improving patient outcomes.

Integrated multicenter deep learning system for prognostic prediction in bladder cancer

Precise survival risk stratification is crucial for personalized therapy in bladder cancer (BCa). This study developed and validated an end-to-end deep learning system using histological slides to predict overall survival (OS) risk in BCa patients. We employed the BlaPaSeg tile classifier to generate tissue probability heatmaps and segmentation maps, trained two prognostic networks, MacroVisionNet and UniVisionNet, and explored six potential BCa prognostic biomarkers. Across all cohorts, the AUC for BlaPaSeg ranged from 0.9906 to 0.9945, while the C-index varied from 0.655 to 0.834 for MacroVisionNet and 0.661 to 0.853 for UniVisionNet. After covariate adjustment, the hazard ratio (HR) values for high-risk groups were 1.97 to 5.06 in MacroVisionNet and 2.13 to 4.01 in UniVisionNet. The high-risk Coloc (Tumor Co-localization score) and IMTS (Integrated Muscle Tumor Score) groups illustrated a higher death risk with HR values from 1.41 to 10.16. The system improves BCa survival prediction and supports refined patient management.

Convolutional Neural Networks (CNNs) for detection of Eye Diseases in an Effective Manner

Introduction: This thesis concentrates on advancing a deep learning system for the precise classification of eye illnesses. Employing the Convolutional Neural Network (CNN) algorithm and Tkinter for GUI, the interface safeguards an intuitive and user-friendly experience. The system effectively identifies various eye conditions, including glaucoma, cataracts, diabetic retinopathy, among others. Objectives: The main goal of this thesis is to offer a reliable and efficient solution for classifying eye diseases through deep learning techniques by addressing the critical need for timely and accurate diagnosis. The significant contributions of this thesis include: Publication of new datasets beneficial to researchers in the field. Enhancement of a robust segmentation approach, applicable to various retinal conditions, including diabetic retinopathy. Methods: A supervised learning approach adopted to train the CNN algorithm with an extensive dataset of eye images. The system achieved an impressive accuracy in classifying eye diseases, underscoring its effectiveness in identification and diagnosis. Development of a hybrid characteristic extraction method from segmented objects, utilized by the classifier for detecting glaucoma and other eye diseases. Results: This system will be able to detect and diagnose the diseases which are related with Eys which will help patients to take care of eyes from further complications. Furthermore, with advancements in portable retinal cameras, integrating the proposed method can facilitate the diagnosis of glaucoma and other eye diseases, thereby enhancing the detection rate, especially in growing nations. Conclusions: The developed system significantly enhances the accuracy and efficiency of eye disease diagnosis, enabling early detection and treatment. The integration of Tkinter for GUI and the color-blind test feature improves the user experience, making the system more accessible to a broader audience. This thesis underscores the potential of deep learning techniques in medical diagnosis, presenting a valuable tool for healthcare professionals and patients alike.

Accurate measurement of key structures in CBD patients using deep learning

The prevalence of common bile duct (CBD) stones presents a widespread and intricate clinical challenge that necessitates precise measurements for effective clinical management. In response, we developed a practical deep learning system designed to precisely measure key structures within endoscopic retrograde cholangiopancreatography (ERCP) images in CBD patients. This system excels in detecting key structures in CBD patients and provides comprehensive information about the quantity, size, location, and dimensions of the common bile duct. While the size, location, and dimensions of the CBD are derived from the scaling ratio of the estimated duodenoscope diameter, the quantity of stones is determined independently of this conversion between pixels and millimeters. A rigorous evaluation of the system’s performance was conducted utilizing diverse metrics such as the Dice coefficient, accuracy, area under the receiver operating characteristic (AUROC) curve, and a comprehensive confusion matrix. A dataset of 418 images from 403 patients was divided into three sets: 314 images (75%) from 299 patients (74%) for training, 52 images (12.5%) from 52 patients (13%) for validation, and 52 images (12.5%) from 52 patients (13%) for testing. The proposed system demonstrated remarkable results in key structure detection and measurement, with Dice coefficients and overall AUROCs of 95.6% and 95.0% for the duodenoscope, 94.6% and 94.0% for the bile duct, and 94.1% and 92.0% for stone identification, respectively. The system’s timely, precise measurements of CBD structures can significantly enhance clinical decision-making, improve patient outcomes, and streamline diagnostic and management processes.

Multi-label semantic segmentation of magnetic resonance images of the prostate gland

Prostate segmentation is a substantial factor in the diagnostic pathway of suspicious prostate lesions. Medical doctors are assisted by computer-aided detection and diagnosis systems systems and methods derived from artificial intelligence deep learning-based systems have to be trained on existing data. Especially freely available labeled prostate magnetic resonance imaging data is rare. With regard to this problem, we show a method to combine two existing small datasets to form a bigger one. We present a data processing pipeline consisting of a cascaded network architecture that is able to perform multi-label semantic segmentation, hence, capable of classifying each pixel in a T2\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ extrm{T}_{2}$$\\end{document}-weighted image not only to one class but to a subset of the prostate zones and classes of interest. This delivers richer information such as overlapping zones that are key to medical radiological examination. Additionally, we describe how to integrate expert knowledge in our deep learning system. To increase data variety for training and evaluation we use image augmentation on our two datasets—a freely available dataset and our new open-source dataset. To combine the datasets we denoise the contourings in our dataset by using an effective yet simple algorithm based on standard computer vision methods only. The performance of the presented methodology is compared and evaluated using the dice score metric and 5-fold cross-validation on all datasets. Although we trained on tiny datasets our method achieves excellent segmentation quality and is even able to detect prostate cancer. Our method to combine the two datasets reduces segmentation errors and increases data variety. The proposed architecture significantly improves performance by including expert knowledge via feature-map concatenation. On the initiative for collaborative computer vision benchmarking dataset we achieve on average dice scores of approximately 91% for the whole prostate gland, 67% for the peripheral zone and 75% for the prostate central gland. We find that image augmentation except contrast limited adaptive histogram equalisation did not have much influence on the segmentation quality. Derived and enhanced from existing methods we present an approach that is able to deliver multi-label semantic segmentation results for prostate magnetic resonance imaging. This approach is simple and could be applied to other applications of deep learning as well. It improves the segmentation results by a large margin. Once tweaked to the data, our denoising and combination algorithm delivers robust and accurate results even on data with segmentation errors.

Development of Pediatric Hypokalemia Prediction Deep-Learning System Based on Wearable Single-Lead Electrocardiography in Real Time

Development of Pediatric Hypokalemia Prediction Deep-Learning System Based on Wearable Single-Lead Electrocardiography in Real Time

A Deep-learning System for Diagnosing Ectopic Eruption

Objectives: To construct a diagnostic model for mixed dentition using a multistage deep-learning network to predict potential ectopic eruption in permanent teeth by integrating dentition segmentation into the process of automatic classification of dental development stages.Methods: A database was established by reviewing 1576 anonymous panoramic radiographs of children aged 6–12 years, collected at the Stomatology Hospital, xxxxxx. These radiographs were categorised as normal or ectopic eruption, with expert diagnoses serving as a benchmark for training and evaluating artificial intelligence (AI) models. Furthermore, tooth boundaries and dental development stages were manually annotated by three pediatric dentistry experts. The dataset was split into training, validation, and test sets at an 8:1:1 ratio.Results: The diagnostic performance of the deep-learning model was rigorously evaluated. The model demonstrated accuracy in tooth segmentation, with Intersection over Union, precision, sensitivity, and F1 scores of 0.959, 0.993, 0.966, and 0.979, respectively. Furthermore, its ability to identify tooth ectopic eruptions on panoramic radiographs, when compared to evaluations by three dentists. Based on McNemar's test, the model's specificity and accuracy in identifying ectopic tooth eruptions on the test dataset surpassed that of Dentist 1 (P < 0.05), while no significant difference was observed compared to the other two dentists. Besides, the deep learning model also showed its potential in classifying dental development stages, as tested against three different standards.Conclusions: The adaptability of the AI-enabled model in this study was demonstrated across multiple scenarios, with clinical validation highlighting its efficacy in diagnosing ectopic eruptions using a multistage deep-learning approach.Clinical Significance: Our findings provide new insights and technical support for the prevention and treatment of abnormal tooth eruption, laying the groundwork for predictive models for other prevalent pediatric dentistry conditions.