Training Labels Research Articles

Closing the Stellar Labels Gap: Stellar Label independent Evidence for [α/M] Information in Gaia BP/RP Spectra

Abstract Data-driven models for stellar spectra that depend on stellar labels suffer from label systematics which decrease model performance: the stellar labels gap. To close the stellar labels gap, we present a stellar label independent model for Gaia BP/RP spectra. We develop a novel implementation of a variational auto-encoder, which learns to generate an XP spectrum and accompanying scatter without relying on stellar labels. We demonstrate that our model achieves competitive XP spectra reconstructions in comparison to stellar label dependent models. We find that our model learns stellar properties directly from the data itself. We then apply our model to XP/APOGEE giant stars to study the [α/M] information in Gaia XP. We provide strong evidence that the XP spectra contain meaningful [α/M] information by demonstrating that our model learns the α-bimodality, without relying on stellar label correlations for stars with T eff < 5000 K, while also being sensitive to the anomalous abundances of Gaia-Enceladus stars. We have publicly released our trained model, codebase and data. Importantly, our stellar label independent model can be implemented for any and all XP spectra because our model's performance scales with training object density, not training label density.

Self-supervised parametric map estimation for multiplexed PET with a deep image prior.

Multiplexed positron emission tomography (mPET) imaging allows simultaneous observation of physiological and pathological information from multiple tracers in a single PET scan. Although supervised deep learning has demonstrated superior performance in mPET image separation compared to purely model-based methods, acquiring large amounts of paired single-tracer data and multi-tracer data for training poses a practical challenge and needs extended scan durations for patients. In addition, the generalisation ability of the supervised learning framework is a concern, as the patient being scanned and their tracer kinetics may potentially fall outside the training distribution. In this work, we propose a self-supervised learning framework based on the deep image prior (DIP) for mPET image separation using just one dataset. In particular, we integrate the multi-tracer compartmental model into the DIP framework to estimate the parametric maps of each tracer from the measured dynamic dual-tracer activity images. Consequently, the separated dynamic single tracer activity images can be recovered from the estimated tracer-specific parametric maps. In the proposed method, dynamic dual-tracer activity images are used as the training label, and the static dual-tracer image (reconstructed from the same patient data from the start to the end of acquisition) is used as the network input. The performance of the proposed method was evaluated on a simulated brain phantom for dynamic dual-tracer [18F]FDG+[11C]MET activity image separation and parametric map estimation. The results demonstrate that the proposed method outperforms the conventional voxel-wise multi-tracer compartmental modeling method (vMTCM) and the two step method DIP-Dn+vMTCM (where dynamic dual-tracer activity images are first denoised using a U-net within the DIP framework, followed by vMTCM separation) in terms of lower bias and standard deviation in the separated single-tracer images and also for the estimated parametric maps for each tracer, at both voxel and ROI levels.

Improving emotion estimation through a combination of ChatGPT and deep learning

ABSTRACT Estimating emotions from text is essential for machines to make human-like decisions, but it remains a challenge due to the significant time and manpower required for labeling training data. Accurate machine learning relies on high-quality training data, often necessitating multiple rounds of evaluation for emotion annotation. Large Language Models (LLMs), such as ChatGPT, hold promise for reducing the labor involved in this process. This study investigates how LLMs can minimize effort while improving accuracy. In Experiment 1, the performance of LLMs in emotion estimation was assessed, showing strong accuracy for negative emotions despite some inconsistencies with human annotations. Experiment 2 introduced a hybrid approach combining LLMs with deep learning techniques to address these inconsistencies, resulting in significantly improved accuracy. These findings demonstrate that integrating LLMs into emotion annotation workflows can reduce manual effort and enhance accuracy, offering a promising pathway for machine learning applications in emotion estimation.

A Deep Learning-Based Watershed Feature Fusion Approach for Tunnel Crack Segmentation in Complex Backgrounds.

As highway tunnel operations continue over time, structural defects, particularly cracks, have been observed to increase annually. Coupled with the rapid expansion of tunnel networks, traditional manual inspection methods have proven inadequate to meet current demands. In recent years, machine vision and deep learning technologies have gained significant attention in civil engineering for the detection and analysis of structural defects. However, rapid and accurate defect identification in highway tunnels presents challenges due to complex background conditions, numerous interfering factors, and the relatively low proportion of cracks within the structure. Additionally, the intensive labor requirements and limited efficiency in labeling training datasets for deep learning pose significant constraints on the deployment of intelligent crack segmentation algorithms. To address these limitations, this study proposes an automatic labeling and optimization algorithm for crack sample sets, utilizing crack features and the watershed algorithm to enable efficient automated segmentation with minimal human input. Furthermore, the deep learning-based crack segmentation network was optimized through comparative analysis of various network depths and residual structure configurations to achieve the best possible model performance. Enhanced accuracy was attained by incorporating axis extraction and watershed filling algorithms to refine segmentation outcomes. Under diverse lining surface conditions and multiple interference factors, the proposed approach achieved a crack segmentation accuracy of 98.78%, with an Intersection over Union (IoU) of 72.41%, providing a robust solution for crack segmentation in tunnels with complex backgrounds.

Fast POI anomaly detection using a weakly-supervised temporal state regression network

Point-of-interest (POI) is a fundamental data type of maps. Anomalous POIs would make maps outdated and lead to user-unfriendly location-based services, and thus should be discovered as fast as possible. Traditional POI anomaly detection methods are inefficient owing to high investigation costs. The emergence of massive human activity data provides a new insight into monitoring POI states through time series modeling. When a POI turns into an anomaly, the associated human activity would disappear. However, human activity data have complicated temporal patterns and noises. It is challenging for existing time series methods to model human activity dynamics. More importantly, there is a lag between the time a POI becomes anomalous and the time we discover it. In this research, we develop a temporal state regression network (TSRNet) model for fast POI anomaly detection. The model can extract temporal features in human activity data, and predict POI state scores as anomaly indicators. Meanwhile, an inference approach is proposed to generate state score sequences as inexact labels for model training. Such weak labels enable TSRNet to identify abnormal temporal patterns as soon as they appear, so that POI outliers can be detected at an early time. Experiments on real-word datasets from AMAP validate the feasibility of our method.

STA-net: a deblurring network combined with spatiotemporal information for zinc froth flotation

Abstract The clarification of blurry frames in froth flotation videos is crucial for the identification and fault diagnosis of zinc flotation conditions. Irregular motion blur represents the primary blur phenomenon in froth flotation videos. Existing methods for motion blur removal have not effectively integrated spatio-temporal information, overlooking the utilization of clear frames, thus failing to achieve satisfactory results in the removal of froth flotation video blur frames. Therefore, this paper proposes a Spatio Temporal Awareness network (STA-net) capable of utilizing the temporal and spatial information from consecutive frames to naturally reconstruct froth images with clarity. This network consists of two sub-networks: the Temporal network (T-net) and the Spatio network (S-net). T-net extracts dynamic information between adjacent clear frames and blurry frames, while S-net captures spatial information within frames and enhances spatial texture details for the coarse images reconstructed based on the temporal information extracted by T-net in a layer-by-layer manner. Additionally, considering that blurry image datasets from actual froth flotation processes often lack corresponding clear images as training labels for the network, a method based on probability density functions (PDF) for generating blurry froth image datasets based clear images is proposed. This method extracts motion vectors from froth images to create blurry froth images that closely resemble the actual blurry conditions of froth images. The deblurring performance of STA-net has been validated in practical froth flotation processes.

Keyphrase Identification Using Minimal Labeled Data with Hierarchical Contexts and Transfer Learning.

Interoperable clinical decision support system (CDSS) rules provide a pathway to interoperability, a well-recognized challenge in health information technology. Building an ontology facilitates creating interoperable CDSS rules, which can be achieved by identifying the keyphrases (KP) from the existing literature. Ontology construction is traditionally a manual effort by human domain experts, and the newly advanced natural language processing techniques, such as KP identification, can be a critical complementary automatic part of building ontology. However, KP identification requires human expertise, consensus, and contextual understanding for data labeling. This paper presents a semi-supervised KP identification framework (long short-term memory-based encoders and the conditional random fields -based decoder models, BiLSTM-CRF) using minimal human labeled data based on hierarchical attention (i.e., at word, sentence, and abstract levels) over the documents and domain adaptation. We created synthetic labels for initial training and human-labeled data for fine-tuning. We also tested different options during NLP preprocessing and ML training to optimize the ML pipeline. Our method outperforms the prior neural architectures by learning through synthetic labels for initial training, document-level contextual learning, language modeling, and fine-tuning with limited gold standard label data. After comparison, we found that the BIO encoding schema performed slightly better than Blue, and domain adaptation techniques can improve the quality of synthetic labels. In addition, document-level context, pre-trained LM, and pre-trained WE all contributed to better model performance in our tasks. Add 2 to 4 human-labeled documents for every 100 synthetic labeled documents improves the model performance without exhausting human-labeled documents too quickly. To the best of our knowledge, this is the first functional framework for the CDSS sub-domain to identify KPs, which is trained on limited human labeled data. It contributes to the general natural language processing (NLP) architectures in areas such as clinical NLP, where manual data labeling is challenging, and light-weighted deep learning models play an important role in real-time KP identification as a complementary approach to human experts' effort.

Detection of Elementary White Mucosal Lesions by an AI System: A Pilot Study

Aim: Accurately identifying primary lesions in oral medicine, particularly elementary white lesions, is a significant challenge, especially for trainee dentists. This study aimed to develop and evaluate a deep learning (DL) model for the detection and classification of elementary white mucosal lesions (EWMLs) using clinical images. Materials and Methods: A dataset was created by collecting photographs of various oral lesions, including oral leukoplakia, OLP plaque-like and reticular forms, OLL, oral candidiasis, and hyperkeratotic lesions from the Unit of Oral Medicine. The SentiSight.AI (Neurotechnology Co.®, Vilnius, Lithuania) AI platform was used for image labeling and model training. The dataset comprised 221 photos, divided into training (n = 179) and validation (n = 42) sets. Results: The model achieved an overall precision of 77.2%, sensitivity of 76.0%, F1 score of 74.4%, and mAP of 82.3%. Specific classes, such as condyloma and papilloma, demonstrated high performance, while others like leucoplakia showed room for improvement. Conclusions: The DL model showed promising results in detecting and classifying EWMLs, with significant potential for educational tools and clinical applications. Expanding the dataset and incorporating diverse image sources are essential for improving model accuracy and generalizability.

Seamless Optimization of Wavelet Parameters for Denoising LFM Radar Signals: An AI-Based Approach

Linear frequency modulation (LFM) signals are pivotal in radar systems, enabling high-resolution measurements and target detection. However, these signals are often degraded by noise, significantly impacting their processing and interpretation. Traditional denoising methods, including wavelet-based techniques, have been extensively used to address this issue, yet they often fall short in terms of optimizing performance due to fixed parameter settings. This paper introduces an innovative approach by combining wavelet denoising with long short-term memory (LSTM) networks specifically tailored for LFM signals in radar systems. By generating a dataset of LFM signals at various signal-to-noise Ratios (SNR) to ensure diversity, we systematically identified the optimal wavelet parameters for each noisy instance. These parameters served as training labels for the proposed LSTM-based architecture, which learned to predict the most effective denoising parameters for a given noisy LFM signal. Our findings reveal a significant enhancement in denoising performance, attributed to the optimized wavelet parameters derived from the LSTM predictions. This advancement not only demonstrates a superior denoising capability but also suggests a substantial improvement in radar signal processing, potentially leading to more accurate and reliable radar detections and measurements. The implications of this paper extend beyond modern radar applications, offering a framework for integrating deep learning techniques with traditional signal processing methods to optimize performance across various noise-dominated domains.

Capturing cell heterogeneity in representations of cell populations for image-based profiling using contrastive learning.

Image-based cell profiling is a powerful tool that compares perturbed cell populations by measuring thousands of single-cell features and summarizing them into profiles. Typically a sample is represented by averaging across cells, but this fails to capture the heterogeneity within cell populations. We introduce CytoSummaryNet: a Deep Sets-based approach that improves mechanism of action prediction by 30-68% in mean average precision compared to average profiling on a public dataset. CytoSummaryNet uses self-supervised contrastive learning in a multiple-instance learning framework, providing an easier-to-apply method for aggregating single-cell feature data than previously published strategies. Interpretability analysis suggests that the model achieves this improvement by downweighting small mitotic cells or those with debris and prioritizing large uncrowded cells. The approach requires only perturbation labels for training, which are readily available in all cell profiling datasets. CytoSummaryNet offers a straightforward post-processing step for single-cell profiles that can significantly boost retrieval performance on image-based profiling datasets.

SsVERDICT: Self-supervised VERDICT-MRI for enhanced prostate tumor characterization.

Demonstrating and assessing self-supervised machine-learning fitting of the VERDICT (vascular, extracellular and restricted diffusion for cytometry in tumors) model for prostate cancer. We derive a self-supervised neural network for fitting VERDICT (ssVERDICT) that estimates parameter maps without training data. We compare the performance of ssVERDICT to two established baseline methods for fitting diffusion MRI models: conventional nonlinear least squares and supervised deep learning. We do this quantitatively on simulated data by comparing the Pearson's correlation coefficient, mean-squared error, bias, and variance with respect to the simulated ground truth. We also calculate in vivo parameter maps on a cohort of 20 prostate cancer patients and compare the methods' performance in discriminating benign from cancerous tissue via Wilcoxon's signed-rank test. In simulations, ssVERDICT outperforms the baseline methods (nonlinear least squares and supervised deep learning) in estimating all the parameters from the VERDICT prostate model in terms of Pearson's correlation coefficient, bias, and mean-squared error. In vivo, ssVERDICT shows stronger lesion conspicuity across all parameter maps, and improves discrimination between benign and cancerous tissue over the baseline methods. ssVERDICT significantly outperforms state-of-the-art methods for VERDICT model fitting and shows, for the first time, fitting of a detailed multicompartment biophysical diffusion MRI model with machine learning without the requirement of explicit training labels.

MTGGAN: generate type-controllable magnetic tile defect data to improve defect detection performance

ABSTRACT The scarcity of training samples and label information poses a major challenge for deep learning defect detection technology. Using generative models to expand sample data is considered an effective approach to address this issue. However, existing GAN models struggle to generate high-fidelity and type-controllable magnetic tile defect samples. Therefore, this paper proposes a type-controllable magnetic tile generation method (MTGGAN) that generates data with inherent label information. The model enhances fidelity and diversity in conditional generation through dual discriminator, auxiliary classifier, and auxiliary generator mechanisms. Additionally, focal loss is introduced to address the model’s class tendency issue during training. Extensive experiments demonstrate that, compared to advanced GAN methods, samples generated by MTGGAN exhibit superior quality and diversity. Furthermore, the generated data can train superior detection models. Effectively alleviating the problem of data scarcity in defect detection.

Computational staining of CD3/CD20 positive lymphocytes in human tissues with experimental confirmation in a genetically engineered mouse model.

Immune dysregulation plays a major role in cancer progression. The quantification of lymphocytic spatial inflammation may enable spatial system biology, improve understanding of therapeutic resistance, and contribute to prognostic imaging biomarkers. In this paper, we propose a knowledge-guided deep learning framework to measure the lymphocytic spatial architecture on human H&E tissue, where the fidelity of training labels is maximized through single-cell resolution image registration of H&E to IHC. We demonstrate that such an approach enables pixel-perfect ground-truth labeling of lymphocytes on H&E as measured by IHC. We then experimentally validate our technique in a genetically engineered, immune-compromised Rag2 mouse model, where Rag2 knockout mice lacking mature lymphocytes are used as a negative experimental control. Such experimental validation moves beyond the classical statistical testing of deep learning models and demonstrates feasibility of more rigorous validation strategies that integrate computational science and basic science. Using our developed approach, we automatically annotated more than 111,000 human nuclei (45,611 CD3/CD20 positive lymphocytes) on H&E images to develop our model, which achieved an AUC of 0.78 and 0.71 on internal hold-out testing data and external testing on an independent dataset, respectively. As a measure of the global spatial architecture of the lymphocytic microenvironment, the average structural similarity between predicted lymphocytic density maps and ground truth lymphocytic density maps was 0.86 ± 0.06 on testing data. On experimental mouse model validation, we measured a lymphocytic density of 96.5 ± %1% in a Rag2+/- control mouse, compared to an average of 16.2 ± %5% in Rag2-/- immune knockout mice (p<0.0001, ANOVA-test). These results demonstrate that CD3/CD20 positive lymphocytes can be accurately detected and characterized on H&E by deep learning and generalized across species. Collectively, these data suggest that our understanding of complex biological systems may benefit from computationally-derived spatial analysis, as well as integration of computational science and basic science.

Automated tree crown labeling with 3D radiative transfer modelling achieves human comparable performances for tree segmentation in semi-arid landscapes

Mapping tree crowns in arid or semi-arid areas, which cover around one-third of the Earth’s land surface, is a key methodology towards sustainable management of trees. Recent advances in deep learning have shown promising results for tree crown segmentation. However, a large amount of manually labeled data is still required. We here propose a novel method to delineate tree crowns from high resolution satellite imagery using deep learning trained with automatically generated labels from 3D radiative transfer modeling, intending to reduce human annotation significantly. The methodological steps consist of 1) simulating images with a 3D radiative transfer model, 2) image style transfer learning based on generative adversarial network (GAN) and 3) tree crown segmentation using U-net segmentation model. The delineation performances of the proposed method have been evaluated on a manually annotated dataset consisting of more than 40,000 tree crowns. Our approach, which relies solely on synthetic images, demonstrates high segmentation accuracy, with an F1 score exceeding 0.77 and an Intersection over Union (IoU) above 0.64. Particularly, it achieves impressive accuracy in extracting crown areas (r2 greater than 0.87) and crown densities (r2 greater than 0.72), comparable to that of a trained dataset with human annotations only. In this study, we demonstrated that the integration of a 3D radiative transfer model and GANs for automatically generating training labels can achieve performances comparable to human labeling, and can significantly reduce the time needed for manual labeling in remote sensing segmentation applications.

Benchmarking machine learning methods for synthetic lethality prediction in cancer

Synthetic lethality (SL) is a gold mine of anticancer drug targets, exposing cancer-specific dependencies of cellular survival. To complement resource-intensive experimental screening, many machine learning methods for SL prediction have emerged recently. However, a comprehensive benchmarking is lacking. This study systematically benchmarks 12 recent machine learning methods for SL prediction, assessing their performance across diverse data splitting scenarios, negative sample ratios, and negative sampling techniques, on both classification and ranking tasks. We observe that all the methods can perform significantly better by improving data quality, e.g., excluding computationally derived SLs from training and sampling negative labels based on gene expression. Among the methods, SLMGAE performs the best. Furthermore, the methods have limitations in realistic scenarios such as cold-start independent tests and context-specific SLs. These results, together with source code and datasets made freely available, provide guidance for selecting suitable methods and developing more powerful techniques for SL virtual screening.

Batch effects correction in scRNA-seq based on biological-noise decoupling autoencoder and central-cross loss

Technical or biologically irrelevant differences caused by different experiments, times, or sequencing platforms can generate batch effects that mask the true biological information. Therefore, batch effects are typically removed when analyzing single-cell RNA sequencing (scRNA-seq) datasets for downstream tasks. Existing batch correction methods usually mitigate batch effects by reducing the data from different batches to a lower dimensional space before clustering, potentially leading to the loss of rare cell types. To address this problem, we introduce a novel single-cell data batch effect correction model using Biological-noise Decoupling Autoencoder (BDA) and Central-cross Loss termed BDACL. The model initially reconstructs raw data using an auto-encoder and conducts preliminary clustering. We then construct a similarity matrix and a hierarchical clustering tree to delineate relationships within and between different batches. Finally, we introduce a Central-cross Loss (CL). This loss leverages cross-entropy loss to prompt the model to better distinguish between different cluster labels. Additionally, it employs the Central Loss to encourage samples to form more compact clusters in the embedding space, thereby enhancing the consistency and interpretability of clustering results to mitigate differences between different batches. The primary innovation of this model lies in reconstructing data with an auto-encoder and gradually merging smaller clusters into larger ones using a hierarchical clustering tree. By using reallocated cluster labels as training labels and employing the Central-cross Loss, the model effectively eliminates batch effects in an unsupervised manner. Compared to current methods, BDACL can mitigate batch effects without losing rare cell types.

Learning Robust Sequential Recommenders through Confident Soft Labels

Sequential recommenders that are trained on implicit feedback are usually learned as a multi-class classification task through softmax-based loss functions on one-hot class labels. However, one-hot training labels are sparse and may lead to biased training and sub-optimal performance. Dense, soft labels have been shown to help improve recommendation performance. However, how to generate high-quality and confident soft labels from noisy sequential interactions between users and items is still an open question. We propose a new learning framework for sequential recommenders, CSRec, which introduces c onfident s oft labels to provide robust guidance when learning from user-item interactions. CSRec contains a teacher module that generates high-quality and confident soft labels and a student module that acts as the target recommender and is trained on the combination of dense, soft labels and sparse, one-hot labels. We propose and compare three approaches to constructing the teacher module: (i) model-level, (ii) data-level, and (iii) training-level. To evaluate the effectiveness and generalization ability of CSRec, we conduct experiments using various state-of-the-art sequential recommendation models as the target student module on four benchmark datasets. Our experimental results demonstrate that CSRec is effective in training better-performing sequential recommenders.

Identifying and Addressing Errors in Anatomical Pathology: A Survey Analysis

Abstract Introduction/Objective To get an update on potential sources of pre-analytical, analytical, post-analytical and other sources of errors in anatomic pathology and strategies to mitigate such errors. Methods/Case Report The authors designed a 5-question survey in Google software and shared it widely on social media platforms. Results (if a Case Study enter NA) The survey, which garnered 43 responses from staff across various areas of the anatomic pathology laboratory, revealed significant insights. Labeling and fixation/preservation errors emerged as the most prevalent pre-analytical errors, while discrepancies between gross and microscopic examinations were common analytical errors. Specimen handling, inadequate clinical information, and transportation were the most frequently perceived other potential sources of errors. The implications of these findings are clear: robust protocols with multiple identifiers, staff training, and barcode labeling can significantly reduce patient identification errors. Similarly, standardized collection instructions, pre-labeled containers, and meticulous verification processes can effectively minimize such errors. Validated tests, standardized scoring systems, and proper investigation of quality control discrepancies are key to reducing analytical errors. Documented procedures, staff training, regular equipment calibration, and reagent testing are also crucial. Conclusion This study contributes to the literature on the current state of potential sources of errors in anatomic pathology. The study findings and recommendations on strategies to address such errors would improve the performance of the anatomic pathology laboratory and eventually improve patient care.

Bootstrap each lead’s latent: A novel method for self-supervised learning of multilead electrocardiograms

Background and Objective:Electrocardiogram (ECG) is one of the most important diagnostic tools for cardiovascular diseases (CVDs). Recent studies show that deep learning models can be trained using labeled ECGs to achieve automatic detection of CVDs, assisting cardiologists in diagnosis. However, the deep learning models heavily rely on labels in training, while manual labeling is costly and time-consuming. This paper proposes a new self-supervised learning (SSL) method for multilead ECGs: bootstrap each lead’s latent (BELL) to reduce the reliance and boost model performance in various tasks, especially when training data are insufficient. Method:BELL is a variant of the well-known bootstrap your own latent (BYOL). The BELL aims to learn prior knowledge from unlabeled ECGs by pretraining, benefitting downstream tasks. It leverages the characteristics of multilead ECGs. First, BELL uses the multiple-branch skeleton, which is more effective in processing multilead ECGs. Moreover, it proposes intra-lead and inter-lead mean square error (MSE) to guide pretraining, and their fusion can result in better performances. Additionally, BELL inherits the main advantage of the BYOL: No negative pair is used in pretraining, making it more efficient. Results:In most cases, BELL surpasses previous works in the experiments. More importantly, the pretraining improves model performances by 0.69% ∼ 8.89% in downstream tasks when only 10% of training data are available. Furthermore, BELL shows excellent adaptability to uncurated ECG data from a real-world hospital. Only slight performance degradation occurs (<1% in most cases) when using these data. Conclusion:The results suggest that the BELL can alleviate the reliance on manual ECG labels from cardiologists, a critical bottleneck of the current deep learning-based models. In this way, the BELL can also help deep learning extend its application on automatic ECG analysis, reducing the cardiologists’ burden in real-world diagnosis.

A dual-labeled dataset and fusion model for automatic teeth segmentation, numbering, and state assessment on panoramic radiographs

BackgroundRecently, deep learning has been increasingly applied in the field of dentistry. The aim of this study is to develop a model for the automatic segmentation, numbering, and state assessment of teeth on panoramic radiographs.MethodsWe created a dual-labeled dataset on panoramic radiographs for training, incorporating both numbering and state labels. We then developed a fusion model that combines a YOLOv9-e instance segmentation model with an EfficientNetv2-l classification model. The instance segmentation model is used for tooth segmentation and numbering, whereas the classification model is used for state evaluation. The final prediction results integrate tooth position, numbering, and state information. The model’s output includes result visualization and automatic report generation.ResultsPrecision, Recall, mAP50 (mean Average Precision), and mAP50-95 for the tooth instance segmentation task are 0.989, 0.955, 0.975, and 0.840, respectively. Precision, Recall, Specificity, and F1 Score for the tooth classification task are 0.943, 0.933, 0.985, and 0.936, respectively.ConclusionsThis fusion model is the first to integrate automatic dental segmentation, numbering, and state assessment. It provides highly accurate results, including detailed visualizations and automated report generation.