Two-stage Classification Research Articles

EXPRESS: From Stars to Dogs – Identifying “Out-Of-Favor” Products on E-Commerce Platforms? Data Analytic Approach to System Design

Online retail platforms are increasingly challenged by the proliferation of low-quality products, which may damage their reputation and sales. To address this problem, we propose a system architecture to proactively identify products that are likely to go "out of favor". Our approach uses historical data to extract useful information from customer ratings and textual reviews. Available data are fed into a state-of-the-art deep learning sequence model to forecast future ratings. We then analyze rating trends, extracting hyperparameters that a binary classifier uses to label products as "out of favor" or not. We tested this system on an Amazon dataset comprising nearly 800,000 observations across 2,826 electronics products. Our results show that the Long Short-Term Memory (LSTM) model excels in forecasting future product ratings compared to other benchmarks. Ablation analysis shows sentiment-related features significantly improve rating forecasts by up to 40%, with review topics adding 10% and other review characteristics 4%. Counterintuitively, topic extraction from reviews does not provide substantial benefits, despite the heavy computational resources it requires. Finally, the two-stage classification process, which leverages time-series data and rating trends, offers a more stable and robust performance than conventional single-stage methods. We provide considerations for system architecture development through robustness checks ensuring its resilience to stressors. Our experiments indicate that rating trends can change in subtle ways over time, leading a promising “star” product to turn into a liability (“dog”). E-commerce platforms can use the proposed system architecture proactively to identify and remove potentially dubious products instead of waiting to take reactive action.

Melon: metagenomic long-read-based taxonomic identification and quantification using marker genes

Long-read sequencing holds great potential for characterizing complex microbial communities, yet taxonomic profiling tools designed specifically for long reads remain lacking. We introduce Melon, a novel marker-based taxonomic profiler that capitalizes on the unique attributes of long reads. Melon employs a two-stage classification scheme to reduce computational time and is equipped with an expectation-maximization-based post-correction module to handle ambiguous reads. Melon achieves superior performance compared to existing tools in both mock and simulated samples. Using wastewater metagenomic samples, we demonstrate the applicability of Melon by showing it provides reliable estimates of overall genome copies, and species-level taxonomic profiles.

Novel Deep CNNs Explore Regions, Boundaries, and Residual Learning for COVID-19 Infection Analysis in Lung CT.

COVID-19 poses a global health crisis, necessitating precise diagnostic methods for timely containment. However, accurately delineating COVID-19-affected regions in lung CT scans is challenging due to contrast variations and significant texture diversity. In this regard, this study introduces a novel two-stage classification and segmentation CNN approach for COVID-19 lung radiological pattern analysis. A novel Residual-BRNet is developed to integrate boundary and regional operations with residual learning, capturing key COVID-19 radiological homogeneous regions, texture variations, and structural contrast patterns in the classification stage. Subsequently, infectious CT images undergo lesion segmentation using the newly proposed RESeg segmentation CNN in the second stage. The RESeg leverages both average and max-pooling implementations to simultaneously learn region homogeneity and boundary-related patterns. Furthermore, novel pixel attention (PA) blocks are integrated into RESeg to effectively address mildly COVID-19-infected regions. The evaluation of the proposed Residual-BRNet CNN in the classification stage demonstrates promising performance metrics, achieving an accuracy of 97.97%, F1-score of 98.01%, sensitivity of 98.42%, and MCC of 96.81%. Meanwhile, PA-RESeg in the segmentation phase achieves an optimal segmentation performance with an IoU score of 98.43% and a dice similarity score of 95.96% of the lesion region. The framework's effectiveness in detecting and segmenting COVID-19 lesions highlights its potential for clinical applications.

Skin Cancer Diagnosis With Multi-Level Classification

Skin cancer arises from the uncontrolled proliferation of abnormal skin cells, primarily triggered by exposure to the harmful ultraviolet (UV) rays of the sun and the utilization of UV tanning beds. This condition poses a heightened risk due to its potential to progress into blood cancer and lead to rapid fatality. Extensive research efforts have been dedicated to advancing the treatment of this perilous ailment. This paper presents a system designed for the examination and diagnosis of pigmented skin lesions and melanoma. The system incorporates a supervised classification algorithm that combines Convolutional Neural Network (CNN) and Deep Neural Network (DNN) architectures with feature extraction techniques. It operates in two distinct stages: the initial stage classifies images into two categories, namely benign or malignant, while the subsequent stage further categorizes the images into one of three classes: basal cell carcinomas, squamous cell carcinomas, or melanoma. Consequently, the comprehensive system addresses four classes, namely benign, basal cell carcinomas, squamous cell carcinomas, and melanoma. This work contributes to the system's design in three significant ways. Firstly, it implements multiple iterations to select the most optimal images, resulting in the highest classification accuracy. Secondly, it employs various statistical methods to identify the most pertinent features, thereby enhancing the classifier's accuracy by focusing on the most informative features for the classification task. Lastly, a two-stage classification approach is implemented, employing two distinct classifiers at different levels within the overall system. Despite the inherent complexity of the real-world problem, the overall system attains a commendable level of classification accuracy. Following rigorous experimentation, the study identifies the top three models. Each approach culminates in a classifier for each stage. The first approach, utilizing a deep learning classifier, achieves an accuracy of 81.82% in the initial cancer discrimination stage and 58.33% in the subsequent stage. The second approach, employing a machine learning classifier, attains an accuracy of 74.63% in the first stage and 64.41% in the second stage. The third approach, utilizing a linear regression classifier, achieves an accuracy of 98% in the first stage and 90% in the second stage. These results underscore the significance of feature selection in influencing model accuracy and suggest the potential for further optimization.

A deep learning-based method for estimating the main stem length of sweet potato seedlings

The length of sweet potato seedling main stems is critical for harvest decisions, but accurately measuring it is challenging due to leaf occlusion and self-curvature. This study proposes a method for stem reconstruction and length estimation using a two-stage classifier, feature extraction, and semantic segmentation to obtain stem masks. Image processing separates clustered skeletons into singular lines for detailed analysis. Object detection identifies roots, stem-leaf intersections, and establishes a search relationship among segments. Stems are reconstructed based on directional, distance, and curvature constraints, then divided into four parts for length estimation using stereo depth information. Experimental results demonstrate effectiveness under various conditions: un-occluded, partially occluded, curved stems, and low-density clusters. Median length estimation error is 1.7 cm, with maximum and minimum errors of 3.2 cm and 0.3 cm respectively. This automated method provides a reliable solution for sweet potato seedling harvest and grading.

The Development of Spinal Endoscopic Ultrasonic Imaging System with an Automated Tissue Recognition Algorithm.

Preclinical experimental study. To develop an intraoperative ultrasound-assisted imaging device, which could be placed at the surgical site through an endoscopic working channel and which could help surgeon recognition of different tissue types during endoscopic spinal surgery (ESS). ESS remains a challenging task for spinal surgeons. Great proficiency and experience are needed to perform procedures such as intervertebral discectomy and neural decompression within a narrow channel. The limited surgical view poses a risk of damaging important structures, such as nerve roots. We constructed a spinal endoscopic ultrasound system, using a 4-mm custom ultrasound probe, which can be easily inserted through the ESS working channel, allowing up to 10 mm depth detection. This system was applied to ovine lumbar spine samples to obtain ultrasound images. Subsequently, we proposed a two-stage classification algorithm, based on a pretrained DenseNet architecture for automated tissue recognition. The recognition algorithm was evaluated using accuracy and consistency. The probe can be easily used in the ESS working channel and produce clear and characteristic ultrasound images. We collected 367 images for training and testing of the recognition algorithm, including images of the spinal cord, nucleus pulposus, adipose tissue, bone, annulus fibrosus and nerve roots. The algorithm achieved over 90% accuracy in recognizing all types of tissues with a Kappa value of 0.875. The recognition times were under 0.1 s using the current configuration. Our system was able to be used in existing ESS working channels and clearly identified at-risk spinal structures in vitro. The pretrained algorithms could identify six intraspinal tissue types accurately and quickly. The concept and innovative application of intraoperative ultrasound in ESS may shorten the learning curve of ESS and improve surgical efficiency and safety.

A Novel Ensemble Deep Learning Based Polyp Detection Using Colonoscopy Dataset

This work addresses the critical task of polyp detection and classification using the SUN colonoscopy video database, which consists of still images annotated with bounding boxes. These images categorize frames into polyp and non-polyp and encompass six distinct classes of polyps: Hyperplastic polyp, Sessile serrated lesion, Low-grade adenoma, Traditional serrated adenoma, High-grade adenoma, and Invasive carcinoma. The approach involves a two-stage classification process. Initially, MobileNetV2 is employed to distinguish between polyp and non-polyp frames. Subsequently, ResNet50 and GoogLeNet are utilized to classify the identified polyps into the six predefined categories. Data augmentation techniques are implemented to address the inherent imbalance in class distribution within the dataset, enhancing model performance and generalizability. The results highlight the effectiveness of GoogLeNet, which achieved an impressive accuracy of 98%, significantly outperforming ResNet50's accuracy of 76.16%. This substantial improvement underscores the potential of GoogLeNet in enhancing the accuracy of polyp classification. The significance of this work lies in its contribution to advancing automated polyp detection and cancer stage classification, crucial for early diagnosis and treatment. These findings provide a foundation for further research and development in this domain, with the potential to improve clinical outcomes through more accurate and timely identification of colorectal polyps.

Two-Stage Classification of Future Knee Osteoarthritis Severity After 8 Years Using MRI: Data from the Osteoarthritis Initiative.

Currently, there are no methods or tools available in clinical practice for classifying future knee osteoarthritis (KOA). In this study, we aimed to fill this gap by classifying future KOA into three severity grades: KL01 (healthy), KL2 (moderate), and KL34 (severe) based on the Kellgren-Lawrance scale. Due to the complex nature of multiclass classification, we used a two-stage method, which separates the classification task into two binary classifications (KL01 vs. KL234 in the first stage and KL2 vs. KL34 in the second stage). Our machine learning (ML) model used two Balanced Random Forest algorithms and was trained with gender, age, height, weight, and quantitative knee morphology obtained from magnetic resonance imaging. Our training dataset comprised longitudinal 8-year follow-up data of 1213 knees from the Osteoarthritis Initiative. Through extensive experimentation with various feature combinations, we identified KL baseline and weight as the most essential features, while gender surprisingly proved to be one of the least influential feature. Our best classification model generated a weighted F1 score of 79.0% and a balanced accuracy of 65.9%. The area under the receiver operating characteristic curve was 83.0% for healthy (KL01) versus moderate (KL2) or severe (KL34) KOA patients and 86.6% for moderate (KL2) versus severe (KL34) KOA patients. We found a statistically significant difference in performance between our two-stage classification model and the traditional single-stage classification model. These findings demonstrate the encouraging results of our two-stage classification model for multiclass KOA severity classification, suggesting its potential application in clinical settings in future.

Deep Learning-Based Intelligent Diagnosis of Lumbar Diseases with Multi-Angle View of Intervertebral Disc

The diagnosis of degenerative lumbar spine disease mainly relies on clinical manifestations and imaging examinations. However, the clinical manifestations are sometimes not obvious, and the diagnosis of medical imaging is usually time-consuming and highly relies on the doctor’s personal experiences. Therefore, a smart diagnostic technology that can assist doctors in manual diagnosis has become particularly urgent. Taking advantage of the development of artificial intelligence, a series of solutions have been proposed for the diagnosis of spinal diseases by using deep learning methods. The proposed methods produce appealing results, but the majority of these approaches are based on sagittal and axial images separately, which limits the capability of different deep learning methods due to the insufficient use of data. In this article, we propose a two-stage classification process that fully utilizes image data. In particular, in the first stage, we used the Mask RCNN model to identify the lumbar spine in the spine image, locate the position of the vertebra and disc, and complete rough classification. In the fine classification stage, a multi-angle view of the intervertebral disc is generated by splicing the sagittal and axial slices of the intervertebral disc up and down based on the key position identified in the first stage, which provides more pieces of information to the deep learning methods for classification. The experimental results reveal substantial performance enhancements with the synthesized multi-angle view, achieving an F1 score of 96.67%. This represents a performance increase of approximately 15% over the sagittal images at 84.48% and nearly 14% over the axial images at 83.15%. This indicates that the proposed paradigm is feasible and more effective in identifying spinal-related degenerative diseases through medical images.

Exploration of MPSO-Two-Stage Classification Optimization Model for Scene Images with Low Quality and Complex Semantics.

Currently, complex scene classification strategies are limited to high-definition image scene sets, and low-quality scene sets are overlooked. Although a few studies have focused on artificially noisy images or specific image sets, none have involved actual low-resolution scene images. Therefore, designing classification models around practicality is of paramount importance. To solve the above problems, this paper proposes a two-stage classification optimization algorithm model based on MPSO, thus achieving high-precision classification of low-quality scene images. Firstly, to verify the rationality of the proposed model, three groups of internationally recognized scene datasets were used to conduct comparative experiments with the proposed model and 21 existing methods. It was found that the proposed model performs better, especially in the 15-scene dataset, with 1.54% higher accuracy than the best existing method ResNet-ELM. Secondly, to prove the necessity of the pre-reconstruction stage of the proposed model, the same classification architecture was used to conduct comparative experiments between the proposed reconstruction method and six existing preprocessing methods on the seven self-built low-quality news scene frames. The results show that the proposed model has a higher improvement rate for outdoor scenes. Finally, to test the application potential of the proposed model in outdoor environments, an adaptive test experiment was conducted on the two self-built scene sets affected by lighting and weather. The results indicate that the proposed model is suitable for weather-affected scene classification, with an average accuracy improvement of 1.42%.

Machine learning-enabled hybrid intrusion detection system with host data transformation and an advanced two-stage classifier

Network Intrusion Detection Systems (NIDS) have been extensively investigated by monitoring real network traffic and analyzing suspicious activities. However, there are limitations in detecting specific types of attacks with NIDS, such as Advanced Persistent Threats (APT). Additionally, NIDS is restricted in observing complete traffic information due to encrypted traffic or a lack of authority. To address these limitations, a Host-based Intrusion Detection system (HIDS) evaluates resources in the host, including logs, files, and folders, to identify APT attacks that routinely inject malicious files into victimized nodes. In this study, a hybrid network intrusion detection system that combines NIDS and HIDS is proposed to improve intrusion detection performance. The host data undergoes a Language Processing (NLP)-based Bidirectional Encoder Representations from Transformers (BERT) model from textual representation to a numerical one in order to process host data in a similar way to the network flow data through machine learning models. The feature flattening technique is applied to flatten two-dimensional host-based features that is provided by BERT into one-dimensional vectors so that host-based and network flow-based features can be processed by advanced Machine Learning (ML) models. In order to enhance HIDS effectiveness, a two-stage collaborative classifier is utilized, which applies two tiers of machine learning algorithms, binary and multi-class classifiers, to detect network intrusions. Once a binary classifier is used to detect benign samples to reduce the complexity of the original problem, the attack data are classified by a multi-class supervised learner to identify attack types. Hence, the overall performance of the two-stage collaborative model outperforms the baseline classifier, XGBoost. The proposed method is shown to generalize across two well-known datasets, CICIDS 2018 and NDSec-1. The performance of XGBoost, which represents conventional ML, is evaluated. Combining host and network features enhances attack detection performance (macro average F1 score) by 8.1% under the CICIDS 2018 dataset and 3.7% under the NDSec-1 dataset. Meanwhile, the two-stage collaborative classifier improves detection performance for most single classes, especially for DoS-LOIC-UDP and DoS-SlowHTTPTest, with improvements of 30.7% and 84.3%, respectively, when compared with the traditional ML models.

A detailed study of resampling algorithms for cyberattack classification in engineering applications.

The evolution of engineering applications is highly relevant in the context of protecting industrial systems. As industries are increasingly interconnected, the need for robust cybersecurity measures becomes paramount. Engineering informatics not only provides tools for knowledge representation and extraction but also affords a comprehensive spectrum of developing sophisticated cybersecurity solutions. However, safeguarding industrial systems poses a unique challenge due to the inherent heterogeneity of data within these environments. Together with this problem, it's crucial to acknowledge that datasets that simulate real cyberattacks within these diverse environments exhibit a high imbalance, often skewed towards certain types of traffics. This study proposes a system for addressing class imbalance in cybersecurity. To do this, three oversampling (SMOTE, Borderline1-SMOTE, and ADASYN) and five undersampling (random undersampling, cluster centroids, NearMiss, repeated edited nearest neighbor, and Tomek Links) methods are tested. Particularly, these balancing algorithms are used to generate one-vs-rest binary models and to develop a two-stage classification system. By doing so, this study aims to enhance the efficacy of cybersecurity measures ensuring a more comprehensive understanding and defense against the diverse range of threats encountered in industrial environments. Experimental results demonstrates the effectiveness of proposed system for cyberattack detection and classification among nine widely known cyberattacks.

Two-Stage Classification Method for Individual Workout Status Prediction with Machine Learning Approach

ABSTRACT The default risk, one of the main risk factors for bonds, should be measured and reflected in the bond yield. Particularly, in the case of financial companies that treat bonds as a major product, failure to properly identify and filter customers’ workout status adversely affects returns. This study proposes a two-stage classification algorithm for workout prediction based on the history data of individual customers such as transaction details of financial companies secured after loans, which is collected over 10 years. The first stage is to rank variables that are closely related to the workout application based on feature selection. In the second step, the first to nth cumulative variables input to each machine learning method generate n candidate classifiers, respectively. Among the total candidates, the model with the highest classification accuracy was selected as the optimal one, which is the Gradient Boost combined with F-score-based feature selection.

A Two-Stage Classification for Dealing with Unseen Clusters in the Testing Data

Classification is an important statistical tool that has increased its importance since the emergence of the data science revolution. However, a training data set that does not capture all underlying population subgroups (or clusters) will result in biased estimates or misclassification. In this paper, we introduce a statistical and computational solution to a possible bias in classification when implemented on estimated population clusters. An unseen-cluster problem denotes the case in which the training data does not contain all underlying clusters in the population. Such a scenario may occur due to various reasons, such as sampling errors, selection bias, or emerging and disappearing population clusters. Once an unseen-cluster problem occurs, a testing observation will be misclassified because a classification rule based on the sample cannot capture a cluster not observed in the training data (sample). To overcome such issues, we suggest a two-stage classification method to ameliorate the unseen-cluster problem in classification. We suggest a test to identify the unseen-cluster problem and demonstrate the performance of the two-stage tailored classifier using simulations and a public data example.

Two-Stage Traffic Sign Classification System

Traffic Sign Recognition (TSR) has received widespread attention due to increased traffic accidents caused by a failure to recognize road traffic signs. Most Traffic Sign Recognition remains challenging due to the illumination of natural scenes, and the diversity in traffic signs [31]. To solve the problem, we proposed a two-stage classification system to overcome the above-mentioned challenges and classify the traffic sign boards efficiently. During the first phase, colour based histogram equalizations applied to eliminates the illumination factors and then traffic sign boards are classified into many shapes, which includes circular, triangular, hexagonal, and diamond traffic signs boards based on the modified HOG feature descriptors and various machine learning classifiers are compared [12]. In the second phase, circular and triangular traffic signs are further categorized into specific subclasses using circular and triangular signs trained Convolutional Neural Network (CNN) respectively to overcome intra-class variation among the similar sign board. Experimentation was conducted on German Traffic Sign Recognition Benchmark dataset (GTSRB) [30] and the results show 99.96% top-1 accuracy during the first stage of classification. In the second stage, there is accuracy of 99.40% for circular traffic signs and 97.50% for triangular traffic signs.

Development of an Anthropomorphic Dental Simulator Based on the Robo-C Robot

An analysis of robots (simulators) in education is provided. Promising directions for their development are highlighted, such as realism, interactivity, adaptation and personalization. The features of using simulators in dentistry are considered. The main disadvantages of existing simulators in dentistry have been identified, namely the lack of a communicative component and imitation of patient behavior. The anthropomorphic dental simulator is based on the Robo-C robot, which is a unique combination of advanced technologies and human facial expressions, which allows it to communicate with people, reproduce movements of different parts of the body and express emotions. As dental components, the following components were created and implemented into the Robo-C control system: a Smart jaw, including cameras and a temperature sensor, and a Smart tooth, including a pressure sensor. The Robo-C control system has been upgraded taking into account the Smart jaw and Smart tooth, which made it possible to connect the dental treatment process with the robot’s servos through its linguistic base. The process of analyzing data obtained from Smart jaw cameras using a neural network is described. A two-stage classification scheme for dental defects has been proposed and its effectiveness has been proven. The linguistic base contains a set of rules with the help of which devices (microphone, speakers, servos, Smart jaw, Smart tooth) interact with each other. An example of compiling a linguistic database rule is given. The linguistic base, Smart-jaw and Smart-tooth are configured for one of four cases: caries treatment, tooth preparation for a crown, tooth extraction, endodontic treatment. Treatment quality control is carried out using a comprehensive assessment of communication interaction with the robot and analysis of Smart-jaw and Smart-tooth data. An example of work in one of the cases is given. The anthropomorphic dental simulator presented in the article allows the use of new technologies in the training of dentists, as well as the simulation of various dental procedures, which will significantly improve the practical preparation of students for working with patients.

A Weighted Ensemble Approach with Multiple Pre-trained Deep Learning Models for Classification of Stroke

Stroke ranks as one of the deadliest diseases globally, emphasizing the crucial need for early diagnosis. This study aims to create a two-stage classification system for stroke and non-stroke images to support early clinical detection. Deep learning, a cornerstone of diagnosis, detection, and prompt treatment, is the primary methodology. Transfer learning adapts successful deep learning architectures for diverse problems, and ensemble learning combines multiple classifiers for enhanced results. These two techniques are applied to classify stroke using a dataset of stroke and normal images. In the initial stage, six pre-trained models are fine-tuned, with DenseNet, Xception, and EfficientNetB2 emerging as the top performers, achieving validation accuracies of 98.4%, 98.4%, and 98%, respectively. These models serve as base learners within an ensemble framework. A weighted average ensemble method combines them, resulting in a remarkable 99.84% accuracy on a reserved test dataset. This approach exhibits promise for stroke detection, a life-threatening condition, while also demonstrating the effectiveness of ensemble techniques in enhancing model performance.

A Four-Stage Mahalanobis-Distance-Based Method for Hand Posture Recognition

Automatic recognition of hand postures is an important research topic with many applications, e.g., communication support for deaf people. In this paper, we present a novel four-stage, Mahalanobis-distance-based method for hand posture recognition using skeletal data. The proposed method is based on a two-stage classification algorithm with two additional stages related to joint preprocessing (normalization) and a rule-based system, specific to hand shapes that the algorithm is meant to classify. The method achieves superior effectiveness on two benchmark datasets, the first of which was created by us for the purpose of this work, while the second is a well-known and publicly available dataset. The method’s recognition rate measured by leave-one-subject-out cross-validation tests is 94.69% on the first dataset and 97.44% on the second. Experiments, including comparison with other state-of-the-art methods and ablation studies related to classification accuracy and time, confirm the effectiveness of our approach.

Bridge Acceleration Data Cleaning Based on Two-Stage Classification Model with Multiple Feature Fusion

Over the past few decades, rapid economic development has led to the establishment of numerous monitoring systems, resulting in the accumulation of vast amounts of monitoring data. Among these data, dynamic acceleration data stand out prominently. However, the quality of collected acceleration data is often compromised due to factors such as challenging operational environments and sensor malfunctions. This severely hampers the value extracted from the data. Although manual identification and classification of data anomalies are more reliable, they are time consuming and labor intensive. To address the challenge of identifying and classifying anomalies in massive acceleration data, this paper proposes a two-stage model for intelligent data cleaning. Firstly, raw acceleration data are transformed into IPDF and PSD features, and a one-dimensional convolutional neural network is trained to preliminarily identify and classify acceleration data anomalies. Subsequently, the RPV indicator is extracted from the original data of the normal and outlier categories to achieve precise classification based on threshold values. The proposed method is successfully validated using acceleration monitoring data from a large-span arch bridge, achieving an accuracy of over 99%. Furthermore, compared to directly employing a one-dimensional CNN classification model, the approach significantly enhances the model’s perception of local significant disturbances.

Classification of Clavicle Fractures based on Multi-View Fusion

Clavicle fracture is a common shoulder injury. Clinical Allman classification divides clavicle fracture into middle fracture, distal fracture and proximal fracture. Different fracture types have corresponding treatment methods and different healing standards. The diagnosis of clavicle fractures can be misdiagnosed and missed by doctors due to blurring of the fracture line. In order to improve the diagnostic efficiency of clinicians and provide clearer treatment ideas, this paper establishes a two-stage clavicle-assisted diagnostic model. The first stage is based on 3D U-Net to segment the shoulder CT of normal clavicle and clavicle fracture in 3D, with dice coefficient reaching 0.9441, and then calculates the two-dimensional image information entropy of the image to select the key layers of the clavicle for classification of the segmented 3D image. The second stage of classification was performed to fuse the key layers data under the three views. The experimental results showed that the three-view fusion had a higher classification accuracy compared to the single-view slice, and the accuracy was improved by 1.3% to 93.4% compared to the best coronal classification, which showed that the two-stage classification method showed good classification effect and could help doctors improve the diagnostic efficiency.