Pre-trained Deep Learning Network Research Articles

CHiMP: deep-learning tools trained on protein crystallization micrographs to enable automation of experiments

A group of three deep-learning tools, referred to collectively as CHiMP (Crystal Hits in My Plate), were created for analysis of micrographs of protein crystallization experiments at the Diamond Light Source (DLS) synchrotron, UK. The first tool, a classification network, assigns images into categories relating to experimental outcomes. The other two tools are networks that perform both object detection and instance segmentation, resulting in masks of individual crystals in the first case and masks of crystallization droplets in addition to crystals in the second case, allowing the positions and sizes of these entities to be recorded. The creation of these tools used transfer learning, where weights from a pre-trained deep-learning network were used as a starting point and repurposed by further training on a relatively small set of data. Two of the tools are now integrated at the VMXi macromolecular crystallography beamline at DLS, where they have the potential to absolve the need for any user input, both for monitoring crystallization experiments and for triggering in situ data collections. The third is being integrated into the XChem fragment-based drug-discovery screening platform, also at DLS, to allow the automatic targeting of acoustic compound dispensing into crystallization droplets.

A Survey on Classifying Ocular Diseases Using Deep Learning and Machine Learning Techniques

In ophthalmology, using fundus images to identify ocular diseases early may pose challenges for clinicians. Manually diagnosing ocular conditions is time-consuming, difficult, and requires experimentation. As a result, technology was created to help computers differentiate between ocular diseases. It is possible to create a system of this type due to different learning algorithms based on visual capabilities. Recent breakthroughs in deep learning and machine learning have led to the development of intelligent systems that improve accuracy and efficiency in classifying eye diseases. The purpose of this study is to conduct a comprehensive survey of modern systems that classify ocular problems using different methods, including pre-trained deep learning networks, using the Ocular Disease Intelligent Recognition (ODIR) dataset. The goal is to build and train a model that can recognize and classify ocular disorders. Previous research indicates the increasing use of deep learning techniques. CNN-based methods have spread widely in this field, compared to traditional manual procedures, due to their outstanding results. The most prominent deep learning techniques are convolutional neural networks (CNNs), recurrent neural networks (RNNs), and various learning methods for increasing and transferring data. The survey highlights the potential of these systems to enhance classification accuracy and sensitivity while addressing challenges such as data availability, interpretability, and integration with clinical practice.

Unlocking the potential: T1-weighed MRI as a powerful predictor of levodopa response in Parkinson’s disease

BackgroundThe efficacy of levodopa, the most crucial metric for Parkinson’s disease diagnosis and treatment, is traditionally gauged through the levodopa challenge test, which lacks a predictive model. This study aims to probe the predictive power of T1-weighted MRI, the most accessible modality for levodopa response.MethodsThis retrospective study used two datasets: from the Parkinson’s Progression Markers Initiative (219 records) and the external clinical dataset from Ruijin Hospital (217 records). A novel feature extraction method using MedicalNet, a pre-trained deep learning network, along with three previous approaches was applied. Three machine learning models were trained and tested on the PPMI dataset and included clinical features, imaging features, and their union set, using the area under the curve (AUC) as the metric. The most significant brain regions were visualized. The external clinical dataset was further evaluated using trained models. A paired one-tailed t-test was performed between the two sets; statistical significance was set at p < 0.001.ResultsFor 46 test set records (mean age, 62 ± 9 years, 28 men), MedicalNet-extracted features demonstrated a consistent improvement in all three machine learning models (SVM 0.83 ± 0.01 versus 0.73 ± 0.01, XgBoost 0.80 ± 0.04 versus 0.74 ± 0.02, MLP 0.80 ± 0.03 versus 0.70 ± 0.07, p < 0.001). Both feature sets were validated on the clinical dataset using SVM, where MedicalNet features alone achieved an AUC of 0.64 ± 0.03. Key responsible brain regions were visualized.ConclusionThe T1-weighed MRI features were more robust and generalizable than the clinical features in prediction; their combination provided the best results. T1-weighed MRI provided insights on specific regions responsible for levodopa response prediction.Critical relevance statementThis study demonstrated that T1w MRI features extracted by a deep learning model have the potential to predict the levodopa response of PD patients and are more robust than widely used clinical information, which might help in determining treatment strategy.Key PointsThis study investigated the predictive value of T1w features for levodopa response.MedicalNet extractor outperformed all other previously published methods with key region visualization.T1w features are more effective than clinical information in levodopa response prediction.Graphical

Human Activity Recognition Algorithm with Physiological and Inertial Signals Fusion: Photoplethysmography, Electrodermal Activity, and Accelerometry.

Inertial signals are the most widely used signals in human activity recognition (HAR) applications, and extensive research has been performed on developing HAR classifiers using accelerometer and gyroscope data. This study aimed to investigate the potential enhancement of HAR models through the fusion of biological signals with inertial signals. The classification of eight common low-, medium-, and high-intensity activities was assessed using machine learning (ML) algorithms, trained on accelerometer (ACC), blood volume pulse (BVP), and electrodermal activity (EDA) data obtained from a wrist-worn sensor. Two types of ML algorithms were employed: a random forest (RF) trained on features; and a pre-trained deep learning (DL) network (ResNet-18) trained on spectrogram images. Evaluation was conducted on both individual activities and more generalized activity groups, based on similar intensity. Results indicated that RF classifiers outperformed corresponding DL classifiers at both individual and grouped levels. However, the fusion of EDA and BVP signals with ACC data improved DL classifier performance compared to a baseline DL model with ACC-only data. The best performance was achieved by a classifier trained on a combination of ACC, EDA, and BVP images, yielding F1-scores of 69 and 87 for individual and grouped activity classifications, respectively. For DL models trained with additional biological signals, almost all individual activity classifications showed improvement (p-value < 0.05). In grouped activity classifications, DL model performance was enhanced for low- and medium-intensity activities. Exploring the classification of two specific activities, ascending/descending stairs and cycling, revealed significantly improved results using a DL model trained on combined ACC, BVP, and EDA spectrogram images (p-value < 0.05).

Elexacaftor/tezacaftor/ivacaftor influences body composition in adults with cystic fibrosis: a fully automated CT-based analysis

A poor nutritional status is associated with worse pulmonary function and survival in people with cystic fibrosis (pwCF). CF transmembrane conductance regulator modulators can improve pulmonary function and body weight, but more data is needed to evaluate its effects on body composition. In this retrospective study, a pre-trained deep-learning network was used to perform a fully automated body composition analysis on chest CTs from 66 adult pwCF before and after receiving elexacaftor/tezacaftor/ivacaftor (ETI) therapy. Muscle and adipose tissues were quantified and divided by bone volume to obtain body size-adjusted ratios. After receiving ETI therapy, marked increases were observed in all adipose tissue ratios among pwCF, including the total adipose tissue ratio (+ 46.21%, p < 0.001). In contrast, only small, but statistically significant increases of the muscle ratio were measured in the overall study population (+ 1.63%, p = 0.008). Study participants who were initially categorized as underweight experienced more pronounced effects on total adipose tissue ratio (p = 0.002), while gains in muscle ratio were equally distributed across BMI categories (p = 0.832). Our findings suggest that ETI therapy primarily affects adipose tissues, not muscle tissue, in adults with CF. These effects are primarily observed among pwCF who were initially underweight. Our findings may have implications for the future nutritional management of pwCF.

Speech Emotion Recognition Using Deep Learning Transfer Models and Explainable Techniques

This study aims to establish a greater reliability compared to conventional speech emotion recognition (SER) studies. This is achieved through preprocessing techniques that reduce uncertainty elements, models that combine the structural features of each model, and the application of various explanatory techniques. The ability to interpret can be made more accurate by reducing uncertain learning data, applying data in different environments, and applying techniques that explain the reasoning behind the results. We designed a generalized model using three different datasets, and each speech was converted into a spectrogram image through STFT preprocessing. The spectrogram was divided into the time domain with overlapping to match the input size of the model. Each divided section is expressed as a Gaussian distribution, and the quality of the data is investigated by the correlation coefficient between distributions. As a result, the scale of the data is reduced, and uncertainty is minimized. VGGish and YAMNet are the most representative pretrained deep learning networks frequently used in conjunction with speech processing. In dealing with speech signal processing, it is frequently advantageous to use these pretrained models synergistically rather than exclusively, resulting in the construction of ensemble deep networks. And finally, various explainable models (Grad CAM, LIME, occlusion sensitivity) are used in analyzing classified results. The model exhibits adaptability to voices in various environments, yielding a classification accuracy of 87%, surpassing that of individual models. Additionally, output results are confirmed by an explainable model to extract essential emotional areas, converted into audio files for auditory analysis using Grad CAM in the time domain. Through this study, we enhance the uncertainty of activation areas that are generated by Grad CAM. We achieve this by applying the interpretable ability from previous studies, along with effective preprocessing and fusion models. We can analyze it from a more diverse perspective through other explainable techniques.

Automated Cough Sound Analysis for Detecting Childhood Pneumonia.

Pneumonia is one of the leading causes of death in children. Prompt diagnosis and treatment can help prevent these deaths, particularly in resource poor regions where deaths due to pneumonia are highest. Clinical symptom-based screening of childhood pneumonia yields excessive false positives, highlighting the necessity for additional rapid diagnostic tests. Cough is a prevalent symptom of acute respiratory illnesses and the sound of a cough can indicate the underlying pathological changes resulting from respiratory infections. In this study, we propose a fully automated approach to evaluate cough sounds to distinguish pneumonia from other acute respiratory diseases in children. The proposed method involves cough sound denoising, cough sound segmentation, and cough sound classification. The denoising algorithm utilizes multi-conditional spectral mapping with a multilayer perceptron network while the segmentation algorithm detects cough sounds directly from the denoised audio waveform. From the segmented cough signal, we extract various handcrafted features and feature embeddings from a pretrained deep learning network. A multilayer perceptron is trained on the combined feature set for detecting pneumonia. The method we propose is evaluated using a dataset comprising cough sounds from 173 children diagnosed with either pneumonia or other acute respiratory diseases. On average, the denoising algorithm improved the signal-to-noise ratio by 44%. Furthermore, a sensitivity and specificity of 91% and 86%, respectively, is achieved in cough segmentation and 82% and 71%, respectively, in detecting childhood pneumonia using cough sounds alone. This demonstrates its potential as a rapid diagnostic tool, such as using smartphone technology.

CLASSIFIED COVID-19 BY DENSENET121-BASED DEEP TRANSFER LEARNING FROM CT-SCAN IMAGES

The COVID-19 disease, which has recently emerged and has been considered a worldwide pandemic, has had a significant impact on the lives of millions of people and has forced a substantial load on healthcare organizations. Numerous deep-learning models have been utilized for diagnosing coronaviruses from chest computed tomography (CT) images. However, in light of the limited availability of datasets on COVID-19, the pre-trained deep learning networks were used. The main objective of this research is to construct and develop an automated approach for the early detection and diagnosis of COVID-19 in thoracic CT images. This paper proposes the DDTL-COV model, a deep transfer learning model based on DenseNet121, to classify patients on CT scans as either COVID or non-COVID, utilizing weights obtained from the ImageNet dataset. Two datasets were used to train the DDTL-COV model: the SARS-CoV-2 CT-scan dataset and the COVID19-CT dataset. In the SARS-CoV-2 CT dataset, the model achieved a good accuracy of 99.6%. However, on the second dataset (COVID19-CT dataset), its performance shows an accuracy rate of 89%. These results show that the model performed better than alternative methods.

To understand double descent, we need to understand VC theory

We analyze generalization performance of over-parameterized learning methods for classification, under VC-theoretical framework. Recently, practitioners in Deep Learning discovered ‘double descent’ phenomenon, when large networks can fit perfectly available training data, and at the same time, achieve good generalization for future (test) data. The current consensus view is that VC-theoretical results cannot account for good generalization performance of Deep Learning networks. In contrast, this paper shows that double descent can be explained by VC-theoretical concepts, such as VC-dimension and Structural Risk Minimization. We also present empirical results showing that double descent generalization curves can be accurately modeled using classical VC-generalization bounds. Proposed VC-theoretical analysis enables better understanding of generalization curves for data sets with different statistical characteristics, such as low vs high-dimensional data and noisy data. In addition, we analyze generalization performance of transfer learning using pre-trained Deep Learning networks.

Smart soil image classification system using lightweight convolutional neural network

In the agriculture sector, soil classification plays a significant task, as it helps in soil tillage, crop selection, moisture level estimation, and automation. Conventionally, soil classification is carried out with the help of physical, chemical, and biological characteristics of the geo-referenced and mapped soil. Soil classification by conventional and laboratory methods is time-consuming, high-cost, and requires proficiency. This study presents a quick and cost-effective prediction of soil type by using soil images. A soil image dataset has been created to classify the soil type using images. To create the soil image dataset, 392 soil samples are collected from different agricultural fields in Andhra Pradesh, India. The collected samples are dried and the soil type is identified using a sieve and hydrometer analysis in the laboratory. An imaging setup has been made to capture the images of the dried soil samples using a smartphone camera. The captured images are pre-processed using: RGB extraction, and V extraction from HSV bins, and adaptive histogram are applied to highlight the texture features of the soil images. A novel lightweight convolutional neural network called Light-SoilNet is proposed to classify five soil sample images: sand, clay, loam, loamy sand, and sandy loam. The proposed network is designed to take care of the imbalanced soil image dataset. The proposed network is tested and compared with state-of-the-art lightweight and pre-trained deep learning networks. The proposed Light-SoilNet network architecture has produced an overall accuracy of 97.2% in classifying the soils. The comparison of the results shows the performance of the proposed model using the image and deep learning techniques in classifying the soil types.

Automatic generation of subject-specific finite element models of the spine from magnetic resonance images.

The generation of subject-specific finite element models of the spine is generally a time-consuming process based on computed tomography (CT) images, where scanning exposes subjects to harmful radiation. In this study, a method is presented for the automatic generation of spine finite element models using images from a single magnetic resonance (MR) sequence. The thoracic and lumbar spine of eight adult volunteers was imaged using a 3D multi-echo-gradient-echo sagittal MR sequence. A deep-learning method was used to generate synthetic CT images from the MR images. A pre-trained deep-learning network was used for the automatic segmentation of vertebrae from the synthetic CT images. Another deep-learning network was trained for the automatic segmentation of intervertebral discs from the MR images. The automatic segmentations were validated against manual segmentations for two subjects, one with scoliosis, and another with a spine implant. A template mesh of the spine was registered to the segmentations in three steps using a Bayesian coherent point drift algorithm. First, rigid registration was applied on the complete spine. Second, non-rigid registration was used for the individual discs and vertebrae. Third, the complete spine was non-rigidly registered to the individually registered discs and vertebrae. Comparison of the automatic and manual segmentations led to dice-scores of 0.93-0.96 for all vertebrae and discs. The lowest dice-score was in the disc at the height of the implant where artifacts led to under-segmentation. The mean distance between the morphed meshes and the segmentations was below 1mm. In conclusion, the presented method can be used to automatically generate accurate subject-specific spine models.

Deep Learning-Based Automatic Diagnosis of Keratoconus with Corneal Endothelium Image.

The primary objective of this study was to develop an end-to-end model that can accurately identify corneal endothelial cells and diagnose keratoconus based on corneal endothelial images acquired from a non-contact specular microscope. This was a retrospective case-control study performed at the Refractive Surgery Center of West China Hospital. A total of 403 keratoconus eyes (221 patients) and 370 myopic eyes (185 normal controls) were consecutively recruited from January 2021 to September 2022. Specular microscopy was used to image and measure the morphometric parameters of the corneal endothelial cells. A Fully Convolutional Network model with a ResNet50 (FCN_ResNet50) was established to perform the endothelial segmentation. The images were then classified using an ensemble machine learning system consisting of four pre-trained deep learning networks: DenseNet121, ResNet50, Inception_v3, and MobileNet_v2. The performance of the models was evaluated based on different metrics, such as accuracy, intersection over union (IoU), and mean IoU. We established a fully end-to-end deep-learning model for the segmentation of endothelial and diagnosis of keratoconus. For endothelial segmentation, the accuracy of the FCN_ResNet50 model achieved near 90% with mean IoU converging to about 80%. The ensemble machine learning system can achieve over 92% accuracy, and > 98% area under curve (AUC) values to diagnose keratoconus with endothelial cell images. In addition, we constructed a diagnostic model based on deep-learning features and developed an associated nomogram which manifested an excellent performance for diagnosis and monitoring the progression of keratoconus. Our research developed an end-to-end model to automatically identify and assess corneal endothelial morphological changes in keratoconus eyes. Moreover, we also constructed a novel nomogram, which can provide valuable information for the diagnosis, monitoring, and management of the disease.

Weightless Neural Network-Based Detection and Diagnosis of Visual Faults in Photovoltaic Modules

The present study introduces a novel approach employing weightless neural networks (WNN) for the detection and diagnosis of visual faults in photovoltaic (PV) modules. WNN leverages random access memory (RAM) devices to simulate the functionality of neurons. The network is trained using a flexible and efficient algorithm designed to produce consistent and precise outputs. The primary advantage of adopting WNN lies in its capacity to obviate the need for network retraining and residual generation, making it highly promising in classification and pattern recognition domains. In this study, visible faults in PV modules were captured using an unmanned aerial vehicle (UAV) equipped with a digital camera capable of capturing RGB images. The collected images underwent preprocessing and resizing before being fed as input into a pre-trained deep learning network, specifically, DenseNet-201, which performed feature extraction. Subsequently, a decision tree algorithm (J48) was employed to select the most significant features for classification. The selected features were divided into training and testing datasets that were further utilized to determine the training, test and validation accuracies of the WNN (WiSARD classifier). Hyperparameter tuning enhances WNN’s performance by achieving optimal values, maximizing classification accuracy while minimizing computational time. The obtained results indicate that the WiSARD classifier achieved a classification accuracy of 100.00% within a testing time of 1.44 s, utilizing the optimal hyperparameter settings. This study underscores the potential of WNN in efficiently and accurately diagnosing visual faults in PV modules, with implications for enhancing the reliability and performance of photovoltaic systems.

Detecting Childhood Pneumonia Using Handcrafted and Deep Learning Cough Sound Features and Multilayer Perceptron.

Pneumonia is one of the leading causes of morbidity and mortality in children. This is especially true in resource poor regions lacking diagnostic facilities, bringing about the need for rapid diagnostic tests for pneumonia. Cough is a common symptom of acute respiratory diseases, including pneumonia, and the sound of cough can be indicative of the pathological variations caused by respiratory infections. As such, in this paper we study objective cough sound evaluation for differentiating between pneumonia and other acute respiratory diseases. We use a dataset of 491 cough sounds from 173 children diagnosed either as having pneumonia or other acute respiratory diseases. We extract features which describe the temporal, spectral, and cepstral characteristics of the cough sound. These features are combined with feature embeddings from a pretrained deep learning network and used to train a multilayer perceptron for classification. The proposed method achieves a sensitivity and specificity of 84% and 73% respectively in differentiating between pneumonia and other acute respiratory diseases using cough sounds alone.

Vision-based toddler physical activity recognition using deep learning

Human activity recognition (HAR) is a system for understanding human movements and behaviour. It has been applied in many fields such as video surveillance, behaviour analysis, and human-computer interaction. The stateof- the-art studies on HAR generally focus their attention on public dataset which mostly consist of adults as their subjects. Research on HAR for children especially toddlers is important to facilitate their surveillance by monitoring their activities automatically. Since toddlers possess different anatomical proportions than adults, their unusual movements can be a challenge to infer. In this paper, a vision-based deep learning HAR system for toddlers was developed based on skeleton features. Videos of toddlers&apos; activities in a day-care were obtained through different public sources. 2D skeleton data were then extracted from every frame of these videos using a pre-trained deep learning network. These skeleton data were trained on LSTM and fully connected network to infer the toddler&apos;s activities. Results showed that this proposed framework managed to achieve 75% accuracies for three toddlers&apos; activities which are jumping, sitting, and standing.

A Novel Shape Retrieval Method for 3D Mechanical Components Based on Object Projection, Pre-Trained Deep Learning Models and Autoencoder

The reuse of existing design models offers great potential in saving resources and generating an efficient workflow. In order to fully benefit from these advantages, it is necessary to develop methods that are able to retrieve mechanical engineering geometry from a query input. This paper aims to address this problem by presenting a method that focuses on the needs of product development to retrieve similar components by comparing the geometrical similarity of existing parts. Therefore, a method is described, which first converts surface meshes into point clouds, rotates them, and then transforms the results into matrices. These are subsequently passed to a pre-trained Deep Learning network to extract the feature vector. A similarity between different geometries is calculated and evaluated based on this vector. The procedure employs a new type of part alignment, especially developed for mechanical engineering geometries. The method is presented in detail and several parameters affecting the accuracy of the retrieval are discussed. This is followed by a critical comparison with other shape retrieval approaches through a mechanical engineering benchmark dataset. • Method for retrieval of mechanical engineering components. • Transformation of geometries with the projection method into matrices. • Application of pre-trained Deep Learning networks to generate the feature vector. • Improving retrieval results through new alignment method. • Comparison of the new procedure with state of the art methods.

Fruit Quality Classification using Convolutional Neural Network

Fruit quality identification is very important in the food industry for maintaining product quality. The quality control in the food industry commonly conducted by human senses which is lack of objectivity and takes long time for real-time mass production quality control. The quality of the fruit can be identified through its color, smell, and texture. This study uses fruit image to classify the quality of the fruit. We trained artificial neural networks for classifying fruit quality from Indian Fruit Dataset with Quality (FruitNet). The dataset contains six classes of fruits with three categorical qualities (Good, Bad, and Mixed). The dataset features were extracted using several pre-trained deep learning networks trained on the ImageNet dataset. The convolutional networks for feature extraction used in this study are VGG16, MobileNetV2, EfficientNetB0, and ResNet50. The extracted features are forwarded to neural network for training the dataset. The result shown that f1-score for testing dataset reaches more than 90% except for MobileNetV2. The highest f1-score is obtained from ResNet50 feature extraction which is 95.7%.

Development and validation of a predictive model combining clinical, radiomics, and deep transfer learning features for lymph node metastasis in early gastric cancer.

BackgroundThis study aims to develop and validate a predictive model combining deep transfer learning, radiomics, and clinical features for lymph node metastasis (LNM) in early gastric cancer (EGC).Materials and methodsThis study retrospectively collected 555 patients with EGC, and randomly divided them into two cohorts with a ratio of 7:3 (training cohort, n = 388; internal validation cohort, n = 167). A total of 79 patients with EGC collected from the Second Affiliated Hospital of Soochow University were used as external validation cohort. Pre-trained deep learning networks were used to extract deep transfer learning (DTL) features, and radiomics features were extracted based on hand-crafted features. We employed the Spearman rank correlation test and least absolute shrinkage and selection operator regression for feature selection from the combined features of clinical, radiomics, and DTL features, and then, machine learning classification models including support vector machine, K-nearest neighbor, random decision forests (RF), and XGBoost were trained, and their performance by determining the area under the curve (AUC) were compared.ResultsWe constructed eight pre-trained transfer learning networks and extracted DTL features, respectively. The results showed that 1,048 DTL features extracted based on the pre-trained Resnet152 network combined in the predictive model had the best performance in discriminating the LNM status of EGC, with an AUC of 0.901 (95% CI: 0.847–0.956) and 0.915 (95% CI: 0.850–0.981) in the internal validation and external validation cohorts, respectively.ConclusionWe first utilized comprehensive multidimensional data based on deep transfer learning, radiomics, and clinical features with a good predictive ability for discriminating the LNM status in EGC, which could provide favorable information when choosing therapy options for individuals with EGC.

Evaluation of handcrafted features and learned representations for the classification of arrhythmia and congestive heart failure in ECG

Electrocardiogram (ECG) is considered as an essential diagnostic tool to investigate life-threatening cardiac abnormalities, such as arrhythmia and congestive heart failure. It is observed that the atrial arrhythmias and congestive heart failure are closely related, wherein, one promotes the other and their co-existence can increase the mortality rate. Timely diagnosis of these diseases is essential to prevent sudden cardiac failure. In this work, we employ a two-fold approach to classify arrhythmia, congestive heart failure, and normal sinus rhythm using ECG fragments. First, we use a traditional hand-crafted feature based model which involves extraction of a number of linear and non-linear features from the ECG fragments. The linear features capture the time-varying and scale of variability information, whereas the non-linear features help to extract the hidden complexity and quantify the uncertainty of the non-stationary signal. Second, an automatic feature learning based approach is employed which uses a pre-trained deep learning network to automatically extract the relevant detailed information from the ECG time–frequency representations. Finally, we explored the combined effect of the two approaches to diagnose the arrhythmias and congestive heart failure patterns. Additionally, this study makes novel use of the subject-level ECG classification. This work on Physionet database shows that the proposed combined system gives an accuracy, sensitivity, specificity, and precision of 99.06%, 99.14%, 99.68%, and 99.32%, respectively, which are better than the state-of-the-art systems. For subject-specific experiment, a further improvement in these performance metrics is obtained using the voting procedure.

Automated Caries Screening Using Ensemble Deep Learning on Panoramic Radiographs

Caries prevention is essential for oral hygiene. A fully automated procedure that reduces human labor and human error is needed. This paper presents a fully automated method that segments tooth regions of interest from a panoramic radiograph to diagnose caries. A patient’s panoramic oral radiograph, which can be taken at any dental facility, is first segmented into several segments of individual teeth. Then, informative features are extracted from the teeth using a pre-trained deep learning network such as VGG, Resnet, or Xception. Each extracted feature is learned by a classification model such as random forest, k-nearest neighbor, or support vector machine. The prediction of each classifier model is considered as an individual opinion that contributes to the final diagnosis, which is decided by a majority voting method. The proposed method achieved an accuracy of 93.58%, a sensitivity of 93.91%, and a specificity of 93.33%, making it promising for widespread implementation. The proposed method, which outperforms existing methods in terms of reliability, and can facilitate dental diagnosis and reduce the need for tedious procedures.