RESUMEN

With the rapid advent and abundance of remote sensing data in different modalities, cross-modal retrieval tasks have gained importance in the research community. Cross-modal retrieval belongs to the research paradigm in which the query is of one modality and the retrieved output is of the other modality. In this paper, the remote sensing (RS) data modalities considered are the earth observation optical data (aerial photos) and the corresponding hand-drawn sketches. The main challenge of the cross-modal retrieval research objective for optical remote sensing images and the corresponding sketches is the distribution gap between the shared embedding space of the modalities. Prior attempts to resolve this issue have not yielded satisfactory outcomes regarding accurately retrieving cross-modal sketch-image RS data. The state-of-the-art architectures used conventional convolutional architectures, which focused on local pixel-wise information about the modalities to be retrieved. This limits the interaction between the sketch texture and the corresponding image, making these models susceptible to overfitting datasets with particular scenarios. To circumvent this limitation, we suggest establishing multi-modal correspondence using a novel architecture of the combined self and cross-attention algorithms, SPCA-Net to minimize the modality gap by employing attention mechanisms for the query and other modalities. Efficient cross-modal retrieval is achieved through the suggested attention architecture, which empirically emphasizes the global information of the relevant query modality and bridges the domain gap through a unique pairwise cross-attention network. In addition to the novel architecture, this paper introduces a unique loss function, label-specific supervised contrastive loss, tailored to the intricacies of the task and to enhance the discriminative power of the learned embeddings. Extensive evaluations are conducted on two sketch-image remote sensing datasets, Earth-on-Canvas and RSketch. Under the same experimental conditions, the performance metrics of our proposed model beat the state-of-the-art architectures by significant margins of 16.7%, 18.9%, 33.7%, and 40.9% correspondingly.

RESUMEN

BACKGROUND: Skin cancer is one of the highly occurring diseases in human life. Early detection and treatment are the prime and necessary points to reduce the malignancy of infections. Deep learning techniques are supplementary tools to assist clinical experts in detecting and localizing skin lesions. Vision transformers (ViT) based on image segmentation classification using multiple classes provide fairly accurate detection and are gaining more popularity due to legitimate multiclass prediction capabilities. MATERIALS AND METHODS: In this research, we propose a new ViT Gradient-Weighted Class Activation Mapping (GradCAM) based architecture named ViT-GradCAM for detecting and classifying skin lesions by spreading ratio on the lesion's surface area. The proposed system is trained and validated using a HAM 10000 dataset by studying seven skin lesions. The database comprises 10 015 dermatoscopic images of varied sizes. The data preprocessing and data augmentation techniques are applied to overcome the class imbalance issues and improve the model's performance. RESULT: The proposed algorithm is based on ViT models that classify the dermatoscopic images into seven classes with an accuracy of 97.28%, precision of 98.51, recall of 95.2%, and an F1 score of 94.6, respectively. The proposed ViT-GradCAM obtains better and more accurate detection and classification than other state-of-the-art deep learning-based skin lesion detection models. The architecture of ViT-GradCAM is extensively visualized to highlight the actual pixels in essential regions associated with skin-specific pathologies. CONCLUSION: This research proposes an alternate solution to overcome the challenges of detecting and classifying skin lesions using ViTs and GradCAM, which play a significant role in detecting and classifying skin lesions accurately rather than relying solely on deep learning models.

Asunto(s)

RESUMEN

This study aimed to design a simple and efficient automatic segmentation model for medical images, so as to facilitate doctors to make more accurate diagnosis and treatment plan. A hybrid lightweight network LIT-Unet with symmetric encoder-decoder U-shaped architecture is proposed. Synapse multi-organ segmentation dataset and automated cardiac diagnosis challenge (ACDC) dataset were used to test the segmentation performance of the method. Two indexes, Dice similarity coefficient (DSC ↑) and 95% Hausdorff distance (HD95 ↓), were used to evaluate and compare the segmentation ability with the current advanced methods. Ablation experiments were conducted to demonstrate the lightweight nature and effectiveness of our model. For Synapse dataset, our model achieves a higher DSC score (80.40%), an improvement of 3.8% over the typical hybrid model (TransUnet). The 95 HD value is low at 20.67%. For ACDC dataset, LIT-Unet achieves the optimal average DSC (%) of 91.84 compared with other networks listed. Compared to patch expanding, the DSC of our model is intuitively improved by 1.62% with the help of deformable token merging (DTM). These results show that the proposed hierarchical LIT-Unet can achieve significant accuracy and is expected to provide a reliable basis for clinical diagnosis and treatment.

RESUMEN

With the rapid pace of urbanization and industrialization, the demand for electricity has surged, placing immense pressure on power management systems. Substation DC systems play a crucial role in managing these fluctuations to ensure a stable and reliable power supply. However, existing battery charging and discharging strategies often suffer from inefficiencies, which can negatively impact overall system performance and sustainability. In this study, we introduce a novel approach that leverages artificial intelligence and time series predictive analytics through the dual self-attention network-neural basis expansion analysis for time series (DSAN-N-BEATS) model. This model integrates the self-attention network with the neural basis expansion analysis for time series (N-BEATS) model to accurately capture time-series data and optimize battery management. Our experimental results demonstrate that the DSAN-N-BEATS model significantly enhances battery state prediction accuracy, achieving a 95.84% accuracy rate, and improves charging and discharging efficiency by 20% compared to traditional methods. These improvements contribute to the overall reliability and sustainability of power systems. This research provides innovative methods for optimizing battery strategies, supporting sustainable development in the power industry, and enhancing system stability and reliability.

RESUMEN

Multiclass classification of brain tumors from magnetic resonance (MR) images is challenging due to high inter-class similarities. To this end, convolution neural networks (CNN) have been widely adopted in recent studies. However, conventional CNN architectures fail to capture the small lesion patterns of brain tumors. To tackle this issue, in this paper, we propose a global transformer network dubbed GT-Net for multiclass brain tumor classification. The GT-Net mainly comprises a global transformer module (GTM), which is introduced on the top of a backbone network. A generalized self-attention block (GSB) is proposed to capture the feature inter-dependencies not only across spatial dimension but also channel dimension, thereby facilitating the extraction of the detailed tumor lesion information while ignoring less important information. Further, multiple GSB heads are used in GTM to leverage global feature dependencies. We evaluate our GT-Net on a benchmark dataset by adopting several backbone networks, and the results demonstrate the effectiveness of GTM. Further, comparison with state-of-the-art methods validates the superiority of our model.

RESUMEN

Estimating travel time (ETA) for arbitrary paths is crucial in urban intelligent transportation systems. Previous studies primarily focus on constructing complex feature systems for individual road segments or sub-segments, which fail to effectively model the influence of each road segment on others. To address this issue, we propose an end-to-end model, AttentionTTE. It utilizes a self-attention mechanism to capture global spatial correlations and a recurrent neural network to capture temporal dependencies from local spatial correlations. Additionally, a multi-task learning module integrates global spatial correlations and temporal dependencies to estimate the travel time for both the entire path and each local path. We evaluate our model on a large trajectory dataset, and extensive experimental results demonstrate that AttentionTTE achieves state-of-the-art performance compared to other methods.

RESUMEN

OBJECTIVES: Accurate segmentation of focal cortical dysplasia (FCD) lesions from MR images plays an important role in surgical planning and decision but is still challenging for radiologists and clinicians. In this study, we introduce a novel transformer-based model, designed for the end-to-end segmentation of FCD lesions from multi-channel MR images. METHODS: The core innovation of our proposed model is the integration of a convolutional neural network-based encoder-decoder structure with a multiscale transformer to augment the feature representation of lesions in the global field of view. Transformer pathways, composed of memory- and computation-efficient dual-self-attention modules, leverage feature maps from varying depths of the encoder to discern long-range interdependencies among feature positions and channels, thereby emphasizing areas and channels relevant to lesions. The proposed model was trained and evaluated on a public-open dataset including MR images of 85 patients using both subject-level and voxel-level metrics. RESULTS: Experimental results indicate that our model offers superior performance both quantitatively and qualitatively. It successfully identified lesions in 82.4% of patients, with a low false-positive lesion cluster rate of 0.176 ± 0.381 per patient. Furthermore, the model achieved an average Dice coefficient of 0.410 ± 0.288, outperforming five established methods. CONCLUSION: Integration of the transformer could enhance the feature presentation and segmentation performance of FCD lesions. The proposed model has the potential to serve as a valuable assistive tool for physicians, enabling rapid and accurate identification of FCD lesions. The source code and pre-trained model weights are available at https://github.com/zhangxd0530/MS-DSA-NET . CRITICAL RELEVANCE STATEMENT: This multiscale transformer-based model performs segmentation of focal cortical dysplasia lesions, aiming to help radiologists and clinicians make accurate and efficient preoperative evaluations of focal cortical dysplasia patients from MR images. KEY POINTS: The first transformer-based model was built to explore focal cortical dysplasia lesion segmentation. Integration of global and local features enhances the segmentation performance of lesions. A valuable benchmark for model development and comparative analyses was provided.

RESUMEN

Alzheimer's disease (AD) severely impacts the lives of many patients and their families. Predicting the progression of the disease from the early stage of mild cognitive impairment (MCI) is of substantial value for treatment, medical research and clinical trials. In this paper, we propose a novel dual attention network to classify progressive MCI (pMCI) and stable MCI (sMCI) using both magnetic resonance imaging (MRI) and neurocognitive metadata. A 3D CNN ShuffleNet V2 model is used as the network backbone to extract MRI image features. Then, neurocognitive metadata is used to guide the spatial attention mechanism to steer the model to focus attention on the most discriminative regions of the brain. In contrast to traditional fusion methods, we propose a ViT based self attention fusion mechanism to fuse the neurocognitive metadata with the 3D CNN feature maps. The experimental results show that our proposed model achieves an accuracy, AUC, and sensitivity of 81.34%, 0.874, and 0.85 respectively using 5-fold cross validation evaluation. A comprehensive experimental study shows our proposed approach significantly outperforms all previous methods for MCI progression classification. In addition, an ablation study shows both fusion methods contribute to the high final performance.

RESUMEN

BACKGROUND: Immune checkpoint inhibitor (ICI) has been widely used in the treatment of advanced cancers, but predicting their efficacy remains challenging. Traditional biomarkers are numerous but exhibit heterogeneity within populations. For comprehensively utilizing the ICI-related biomarkers, we aim to conduct multidimensional feature selection and deep learning model construction. METHODS: We used statistical and machine learning methods to map features of different levels to next-generation sequencing gene expression. We integrated genes from different sources into the feature input of a deep learning model, by means of self-attention mechanism. RESULTS: We performed feature selection at the single-cell sequencing level, PD-L1 (CD274) analysis level, tumor mutational burden (TMB)/mismatch repair (MMR) level, and somatic copy number alteration (SCNA) level, obtaining 96 feature genes. Based on the pan-cancer dataset, we trained a multi-task deep learning model. We tested the model in the bladder urothelial carcinoma testing set 1 (AUC = 0.62, n = 298), bladder urothelial carcinoma testing set 2 (AUC = 0.66, n = 89), non-small cell lung cancer testing set (AUC = 0.85, n = 27), and skin cutaneous melanoma testing set (AUC = 0.71, n = 27). CONCLUSION: Our study demonstrates the potential of the deep learning model for integrating multidimensional features in predicting the outcome of ICI. Our study also provides a potential methodological case for medical scenarios requiring the integration of multiple levels of features.

RESUMEN

Combining item feature information helps extract comprehensive sequential patterns, thereby improving the accuracy of sequential recommendations. However, existing methods usually combine features of each item using a vanilla attention mechanism. We argue that such a combination ignores the interactions between features and does not model integrated feature representations. In this study, we propose a novel Feature Interaction Dual Self-attention network (FIDS) model for sequential recommendation, which utilizes dual self-attention to capture both feature interactions and sequential transition patterns. Specifically, we first model the feature interactions for each item to form meaningful higher-order feature representations using a multi-head attention mechanism. Then, we adopt two independent self-attention networks to capture the transition patterns in both the item sequence and the integrated feature sequence, respectively. Moreover, we stack multiple self-attention blocks and add residual connections at each block for all self-attention networks. Finally, we combine the feature-wise and item-wise sequential patterns into a fully connected layer for the next item recommendation. We conduct experiments on two real-world datasets, and our experimental results show that the proposed FIDS method outperforms state-of-the-art recommendation models.

RESUMEN

BACKGROUND: Fall detection is of great significance in safeguarding human health. By monitoring the motion data, a fall detection system (FDS) can detect a fall accident. Recently, wearable sensors-based FDSs have become the mainstream of research, which can be categorized into threshold-based FDSs using experience, machine learning-based FDSs using manual feature extraction, and deep learning (DL)-based FDSs using automatic feature extraction. However, most FDSs focus on the global information of sensor data, neglecting the fact that different segments of the data contribute variably to fall detection. This shortcoming makes it challenging for FDSs to accurately distinguish between similar human motion patterns of actual falls and fall-like actions, leading to a decrease in detection accuracy. OBJECTIVE: This study aims to develop and validate a DL framework to accurately detect falls using acceleration and gyroscope data from wearable sensors. We aim to explore the essential contributing features extracted from sensor data to distinguish falls from activities of daily life. The significance of this study lies in reforming the FDS by designing a weighted feature representation using DL methods to effectively differentiate between fall events and fall-like activities. METHODS: Based on the 3-axis acceleration and gyroscope data, we proposed a new DL architecture, the dual-stream convolutional neural network self-attention (DSCS) model. Unlike previous studies, the used architecture can extract global feature information from acceleration and gyroscope data. Additionally, we incorporated a self-attention module to assign different weights to the original feature vector, enabling the model to learn the contribution effect of the sensor data and enhance classification accuracy. The proposed model was trained and tested on 2 public data sets: SisFall and MobiFall. In addition, 10 participants were recruited to carry out practical validation of the DSCS model. A total of 1700 trials were performed to test the generalization ability of the model. RESULTS: The fall detection accuracy of the DSCS model was 99.32% (recall=99.15%; precision=98.58%) and 99.65% (recall=100%; precision=98.39%) on the test sets of SisFall and MobiFall, respectively. In the ablation experiment, we compared the DSCS model with state-of-the-art machine learning and DL models. On the SisFall data set, the DSCS model achieved the second-best accuracy; on the MobiFall data set, the DSCS model achieved the best accuracy, recall, and precision. In practical validation, the accuracy of the DSCS model was 96.41% (recall=95.12%; specificity=97.55%). CONCLUSIONS: This study demonstrates that the DSCS model can significantly improve the accuracy of fall detection on 2 publicly available data sets and performs robustly in practical validation.

Asunto(s)

RESUMEN

Mine flooding accidents have occurred frequently in recent years, and the predicting of mine water inflow is one of the most crucial flood warning indicators. Further, the mine water inflow is characterized by non-linearity and instability, making it difficult to predict. Accordingly, we propose a time series prediction model based on the fusion of the Transformer algorithm, which relies on self-attention, and the LSTM algorithm, which captures long-term dependencies. In this paper, Baotailong mine water inflow in Heilongjiang Province is used as sample data, and the sample data is divided into different ratios of the training set and test set in order to obtain optimal prediction results. In this study, we demonstrate that the LSTM-Transformer model exhibits the highest training accuracy when the ratio is 7:3. To improve the efficiency of search, the combination of random search and Bayesian optimization is used to determine the network model parameters and regularization parameters. Finally, in order to verify the accuracy of the LSTM-Transformer model, the LSTM-Transformer model is compared with LSTM, CNN, Transformer and CNN-LSTM models. The results prove that LSTM-Transformer has the highest prediction accuracy, and all the indicators of its model are well improved.

RESUMEN

Objective.Automatic segmentation of prostatic zones from MRI can improve clinical diagnosis of prostate cancer as lesions in the peripheral zone (PZ) and central gland (CG) exhibit different characteristics. Existing approaches are limited in their accuracy in localizing the edges of PZ and CG. The proposed boundary-aware semantic clustering network (BASC-Net) improves segmentation performance by learning features in the vicinity of the prostate zonal boundaries, instead of only focusing on manually segmented boundaries.Approach.BASC-Net consists of two major components: the semantic clustering attention (SCA) module and the boundary-aware contrastive (BAC) loss. The SCA module implements a self-attention mechanism that extracts feature bases representing essential features of the inner body and boundary subregions and constructs attention maps highlighting each subregion. SCA is the first self-attention algorithm that utilizes ground truth masks to supervise the feature basis construction process. The features extracted from the inner body and boundary subregions of the same zone were integrated by BAC loss, which promotes the similarity of features extracted in the two subregions of the same zone. The BAC loss further promotes the difference between features extracted from different zones.Main results.BASC-Net was evaluated on the NCI-ISBI 2013 Challenge and Prostate158 datasets. An inter-dataset evaluation was conducted to evaluate the generalizability of the proposed method. BASC-Net outperformed nine state-of-the-art methods in all three experimental settings, attaining Dice similarity coefficients of 79.9% and 88.6% for PZ and CG, respectively, in the NCI-ISBI dataset, 80.5% and 89.2% for PZ and CG, respectively, in Prostate158 dataset, and 73.2% and 87.4% for PZ and CG, respectively, in the inter-dataset evaluation.Significance.As prostate lesions in PZ and CG have different characteristics, the zonal boundaries segmented by BASC-Net will facilitate prostate lesion detection.

Asunto(s)

RESUMEN

Background: Bone age assessment (BAA) is crucial for the diagnosis of growth disorders and the optimization of treatments. However, the random error caused by different observers' experiences and the low consistency of repeated assessments harms the quality of such assessments. Thus, automated assessment methods are needed. Methods: Previous research has sought to design localization modules in a strongly or weakly supervised fashion to aggregate part regions to better recognize subtle differences. Conversely, we sought to efficiently deliver information between multi-granularity regions for fine-grained feature learning and to directly model long-distance relationships for global understanding. The proposed method has been named the "Multi-Granularity and Multi-Attention Net (2M-Net)". Specifically, we first applied the jigsaw method to generate related tasks emphasizing regions with different granularities, and we then trained the model on these tasks using a hierarchical sharing mechanism. In effect, the training signals from the extra tasks created as an inductive bias, enabling 2M-Net to discover task relatedness without the need for annotations. Next, the self-attention mechanism acted as a plug-and-play module to effectively enhance the feature representation capabilities. Finally, multi-scale features were applied for prediction. Results: A public data set of 14,236 hand radiographs, provided by the Radiological Society of North America (RSNA), was used to develop and validate 2M-Net. In the public benchmark testing, the mean absolute error (MAE) between the bone age estimates of the model and of the reviewer was 3.98 months (3.89 months for males and 4.07 months for females). Conclusions: By using the jigsaw method to construct a multi-task learning strategy and inserting the self-attention module for efficient global modeling, we established 2M-Net, which is comparable to the previous best method in terms of performance.

RESUMEN

Background: Axillary lymph node (ALN) status is a crucial prognostic indicator for breast cancer metastasis, with manual interpretation of whole slide images (WSIs) being the current standard practice. However, this method is subjective and time-consuming. Recent advancements in deep learning-based methods for medical image analysis have shown promise in improving clinical diagnosis. This study aims to leverage these technological advancements to develop a deep learning model based on features extracted from primary tumor biopsies for preoperatively identifying ALN metastasis in early-stage breast cancer patients with negative nodes. Methods: We present DLCNBC-SA, a deep learning-based network specifically tailored for core needle biopsy and clinical data feature extraction, which integrates a self-attention mechanism (CNBC-SA). The proposed model consists of a feature extractor based on convolutional neural network (CNN) and an improved self-attention mechanism module, which can preserve the independence of features in WSIs for analysis and enhancement to provide rich feature representation. To validate the performance of the proposed model, we conducted comparative experiments and ablation studies using publicly available datasets, and verification was performed through quantitative analysis. Results: The comparative experiment illustrates the superior performance of the proposed model in the task of binary classification of ALNs, as compared to alternative methods. Our method achieved outstanding performance [area under the curve (AUC): 0.882] in this task, significantly surpassing the state-of-the-art (SOTA) method on the same dataset (AUC: 0.862). The ablation experiment reveals that incorporating RandomRotation data augmentation technology and utilizing Adadelta optimizer can effectively enhance the performance of the proposed model. Conclusions: The experimental results demonstrate that the model proposed in this paper outperforms the SOTA model on the same dataset, thereby establishing its reliability as an assistant for pathologists in analyzing WSIs of breast cancer. Consequently, it significantly enhances both the efficiency and accuracy of doctors during the diagnostic process.

RESUMEN

BACKGROUND: Chest X-ray image classification for multiple diseases is an important research direction in the field of computer vision and medical image processing. It aims to utilize advanced image processing techniques and deep learning algorithms to automatically analyze and identify X-ray images, determining whether specific pathologies or structural abnormalities exist in the images. OBJECTIVE: We present the MMPDenseNet network designed specifically for chest multi-label disease classification. METHODS: Initially, the network employs the adaptive activation function Meta-ACON to enhance feature representation. Subsequently, the network incorporates a multi-head self-attention mechanism, merging the conventional convolutional neural network with the Transformer, thereby bolstering the ability to extract both local and global features. Ultimately, the network integrates a pyramid squeeze attention module to capture spatial information and enrich the feature space. RESULTS: The concluding experiment yielded an average AUC of 0.898, marking an average accuracy improvement of 0.6% over the baseline model. When compared with the original network, the experimental results highlight that MMPDenseNet considerably elevates the classification accuracy of various chest diseases. CONCLUSION: It can be concluded that the network, thus, holds substantial value for clinical applications.

RESUMEN

The manufacturing industry has been operating within a constantly evolving technological environment, underscoring the importance of maintaining the efficiency and reliability of manufacturing processes. Motor-related failures, especially bearing defects, are common and serious issues in manufacturing processes. Bearings provide accurate and smooth movements and play essential roles in mechanical equipment with shafts. Given their importance, bearing failure diagnosis has been extensively studied. However, the imbalance in failure data and the complexity of time series data make diagnosis challenging. Conventional AI models (convolutional neural networks (CNNs), long short-term memory (LSTM), support vector machine (SVM), and extreme gradient boosting (XGBoost)) face limitations in diagnosing such failures. To address this problem, this paper proposes a bearing failure diagnosis model using a graph convolution network (GCN)-based LSTM autoencoder with self-attention. The model was trained on data extracted from the Case Western Reserve University (CWRU) dataset and a fault simulator testbed. The proposed model achieved 97.3% accuracy on the CWRU dataset and 99.9% accuracy on the fault simulator dataset.

RESUMEN

Bacteriophages are the viruses that infect bacterial cells. They are the most diverse biological entities on earth and play important roles in microbiome. According to the phage lifestyle, phages can be divided into the virulent phages and the temperate phages. Classifying virulent and temperate phages is crucial for further understanding of the phage-host interactions. Although there are several methods designed for phage lifestyle classification, they merely either consider sequence features or gene features, leading to low accuracy. A new computational method, DeePhafier, is proposed to improve classification performance on phage lifestyle. Built by several multilayer self-attention neural networks, a global self-attention neural network, and being combined by protein features of the Position Specific Scoring Matrix matrix, DeePhafier improves the classification accuracy and outperforms two benchmark methods. The accuracy of DeePhafier on five-fold cross-validation is as high as 87.54% for sequences with length >2000bp.

Asunto(s)

RESUMEN

BACKGROUND: Growing evidence suggests that distal regulatory elements are essential for cellular function and states. The sequences within these distal elements, especially motifs for transcription factor binding, provide critical information about the underlying regulatory programs. However, cooperativities between transcription factors that recognize these motifs are nonlinear and multiplexed, rendering traditional modeling methods insufficient to capture the underlying mechanisms. Recent development of attention mechanism, which exhibit superior performance in capturing dependencies across input sequences, makes them well-suited to uncover and decipher intricate dependencies between regulatory elements. RESULT: We present Transcription factors cooperativity Inference Analysis with Neural Attention (TIANA), a deep learning framework that focuses on interpretability. In this study, we demonstrated that TIANA could discover biologically relevant insights into co-occurring pairs of transcription factor motifs. Compared with existing tools, TIANA showed superior interpretability and robust performance in identifying putative transcription factor cooperativities from co-occurring motifs. CONCLUSION: Our results suggest that TIANA can be an effective tool to decipher transcription factor cooperativities from distal sequence data. TIANA can be accessed through: https://github.com/rzzli/TIANA .

Asunto(s)

RESUMEN

BACKGROUND: Maxillary expansion is an important treatment method for maxillary transverse hypoplasia. Different methods of maxillary expansion should be carried out depending on the midpalatal suture maturation levels, and the diagnosis was validated by palatal plane cone beam computed tomography (CBCT) images by orthodontists, while such a method suffered from low efficiency and strong subjectivity. This study develops and evaluates an enhanced vision transformer (ViT) to automatically classify CBCT images of midpalatal sutures with different maturation stages. METHODS: In recent years, the use of convolutional neural network (CNN) to classify images of midpalatal suture with different maturation stages has brought positive significance to the decision of the clinical maxillary expansion method. However, CNN cannot adequately learn the long-distance dependencies between images and features, which are also required for global recognition of midpalatal suture CBCT images. The Self-Attention of ViT has the function of capturing the relationship between long-distance pixels of the image. However, it lacks the inductive bias of CNN and needs more data training. To solve this problem, a CNN-enhanced ViT model based on transfer learning is proposed to classify midpalatal suture CBCT images. In this study, 2518 CBCT images of the palate plane are collected, and the images are divided into 1259 images as the training set, 506 images as the verification set, and 753 images as the test set. After the training set image preprocessing, the CNN-enhanced ViT model is trained and adjusted, and the generalization ability of the model is tested on the test set. RESULTS: The classification accuracy of our proposed ViT model is 95.75%, and its Macro-averaging Area under the receiver operating characteristic Curve (AUC) and Micro-averaging AUC are 97.89% and 98.36% respectively on our data test set. The classification accuracy of the best performing CNN model EfficientnetV2_S was 93.76% on our data test set. The classification accuracy of the clinician is 89.10% on our data test set. CONCLUSIONS: The experimental results show that this method can effectively complete CBCT images classification of midpalatal suture maturation stages, and the performance is better than a clinician. Therefore, the model can provide a valuable reference for orthodontists and assist them in making correct a diagnosis.

Asunto(s)