Faster Region-based Convolutional Neural Network Research Articles

Detection of complicated bleeding points based on a faster region-based convolutional neural network.

Detection of complicated bleeding points based on a faster region-based convolutional neural network.

Deep learning and computer vision based occupancy CO2 level prediction for demand-controlled ventilation (DCV)

The present study investigated the potential of the application of a live occupancy detection approach to assist the operations of demand-controlled ventilation (DCV) systems to ensure that sufficient interior thermal conditions and air quality were attained while reducing unnecessary building energy loads to improve building energy performance. Faster region-based convolutional neural network (RCNN) models were trained to detect the number of people and occupancy activities respectively, and deployed to an artificial intelligence (AI)-powered camera. Experimental tests were carried out within a case study room to assess the performance of this approach. Due to the less complexity of people counting model, it achieved an average intersection over union (IoU) detection accuracy of about 98.9%, which was higher than activity detection model of about 88.5%. During the detection, the count-based occupancy profiles were produced according to the real-time information about the number of people and their activities. To estimate the effect of this approach on indoor air quality and energy demand, scenario-based modelling of the case study building under four ventilation scenarios was carried out via building energy simulation (BES). Results showed that the proposed approach could provide demand-driven ventilation controls data on the dynamic changes of occupancy to improve the indoor air quality (IAQ) and address the problem of under- or over-estimation of the ventilation demand when using the static or fixed profiles.

A Few-Shot Dental Object Detection Method Based on a Priori Knowledge Transfer

With the continuous improvement in oral health awareness, people’s demand for oral health diagnosis has also increased. Dental object detection is a key step in automated dental diagnosis; however, because of the particularity of medical data, researchers usually cannot obtain sufficient medical data. Therefore, this study proposes a dental object detection method for small-size datasets based on teeth semantics, structural information feature extraction, and an a priori knowledge migration, called a segmentation, points, segmentation, and classification network (SPSC-NET). In the region of interest area extraction method, the SPSC-NET method converts the teeth X-ray image into an a priori knowledge information image, composed of the edges of the teeth and the semantic segmentation image; the network structure used to extract the a priori knowledge information is a symmetric structure, which then generates the key points of the object instance. Next, it uses the key points of the object instance (i.e., the dental semantic segmentation image and the dental edge image) to obtain the object instance image (i.e., the positioning of the teeth). Using 10 training images, the test precision and recall rate of the tooth object center point of the SPSC-NET method were between 99–100%. In the classification method, the SPSC-NET identified the single instance segmentation image generated by migrating the dental object area, the edge image, and the semantic segmentation image as a priori knowledge. Under the premise of using the same deep neural network classification model, the model classification with a priori knowledge was 20% more accurate than the ordinary classification methods. For the overall object detection performance indicators, the SPSC-NET’s average precision (AP) value was more than 92%, which is better than that of the transfer-based faster region-based convolutional neural network (Faster-RCNN) object detection model; moreover, its AP and mean intersection-over-union (mIOU) were 14.72% and 19.68% better than the transfer-based Faster-CNN model, respectively.

Deep Learning Application in Dental Caries Detection Using Intraoral Photos Taken by Smartphones

A mobile-phone-based diagnostic tool, which most of the population can easily access, could be a game changer in increasing the number of examinations of people with dental caries. This study aimed to apply a deep learning algorithm in diagnosing the stages of smooth surface caries via smartphone images. Materials and methods: A training dataset consisting of 1902 photos of the smooth surface of teeth taken with an iPhone 7 from 695 people was used. Four deep learning models, consisting of Faster Region-Based Convolutional Neural Networks (Faster R-CNNs), You Only Look Once version 3 (YOLOv3), RetinaNet, and Single-Shot Multi-Box Detector (SSD), were tested to detect initial caries lesions and cavities. The reference standard was the diagnosis of a dentist based on image examination according to the International Caries Classification and Management System (ICCMS) classification. Results: For cavitated caries, YOLOv3 and Faster R-CNN showed the highest sensitivity among the four tested models, at 87.4% and 71.4%, respectively. The sensitivity levels of these two models were only 36.9 % and 26% for visually non-cavitated (VNC). The specificity of the four models reached above 86% for cavitated caries and above 71% for VNC. Conclusion: The clinical application of YOLOv3 and Faster R-CNN models for diagnosing dental caries via smartphone images was promising. The current study provides a preliminary insight into the potential translation of AI from the laboratory to clinical practice.

A Deep Learning System to Predict the Histopathological Results From Urine Cytopathological Images.

BackgroundAlthough deep learning systems (DLSs) have been developed to diagnose urine cytology, more evidence is required to prove if such systems can predict histopathology results as well.MethodsWe retrospectively retrieved urine cytology slides and matched histological results. High-power field panel images were annotated by a certified urological pathologist. A deep learning system was designed with a ResNet101 Faster R-CNN (faster region-based convolutional neural network). It was firstly built to spot cancer cells. Then, it was directly used to predict the likelihood of the presence of tissue malignancy.ResultsWe retrieved 441 positive cases and 395 negative cases. The development involved 387 positive cases, accounting for 2,668 labeled cells, to train the DLS to spot cancer cells. The DLS was then used to predict corresponding histopathology results. In an internal test set of 85 cases, the area under the curve (AUC) was 0.90 (95%CI 0.84–0.96), and the kappa score was 0.68 (95%CI 0.52–0.84), indicating substantial agreement. The F1 score was 0.56, sensitivity was 71% (95%CI 52%–85%), and specificity was 94% (95%CI 84%–98%). In an extra test set of 333 cases, the DLS achieved 0.25 false-positive cells per image. The AUC was 0.93 (95%CI 0.90–0.95), and the kappa score was 0.58 (95%CI 0.46–0.70) indicating moderate agreement. The F1 score was 0.66, sensitivity was 67% (95%CI 54%–78%), and specificity was 92% (95%CI 88%–95%).ConclusionsThe deep learning system could predict if there was malignancy using cytocentrifuged urine cytology images. The process was explainable since the prediction of malignancy was directly based on the abnormal cells selected by the model and can be verified by examining those candidate abnormal cells in each image. Thus, this DLS was not just a tool for pathologists in cytology diagnosis. It simultaneously provided novel histopathologic insights for urologists.

Distributed detection of internal cavities in concrete-filled steel tube arch bridge elements

Internal cavities in structural elements, such as in concrete-filled steel tubes (CFSTs) of arch ribs reduce confinement, bearing capacity, and the durability of the arch bridges. Formation of internal cavities may be either materials related, which affects proper consolidation, or construction related, especially in terms of proper timing and delivery of concrete slurry to horizontal and inclined structural members. Nondestructive test methods such as ultrasonics have been employed for detection of internal cavities. Despite their success, point-by-point inspection of structural members becomes time-consuming and inefficient, especially when entire structures need to be inspected. Infrared thermography (IR) provides a more reasonable global approach for detection of anomalies. The accuracy and resolution of IR depend on the local ambient conditions affecting convection and external thermal flux prior to detection by IR. The research described herein pertains to the development of a hybrid thermographic approach by using Brillouin fiber optic sensors (BFOSs) for direct detection of thermal convection at the surfaces of steel tubes. Accurate quantitative detection of internal cavity locations and their dimensions required development of a specific machine learning-based approach through which the distributed thermographic data was converted to a series of grayscale images. A Faster Region-based Convolutional Neural Network (Faster R-CNN) based on the improved VGG-16 (IVGG-16) network architecture was utilized for this purpose. The experimental work involved design and fabrication of a distributed temperature sensing sheet (DTSS) for simultaneous transmission of thermal energy into CFST specimens and acquisition of distributed thermographic data by the BFOS. By using the proposed approach, it was possible to detect the embedded internal cavities during the experiments.

Ship detection of coast defense radar in real marine environment based on fast YOLO V4

We propose an innovative method for ship detection in the real marine environment. Different from other high-resolution optical images and synthetic aperture radar images, ship detection of coast defense radar is quite challenging due to the complex background, sea state, and low resolution. To this end, we build a real dataset and propose an innovative detection method based on you only look once (YOLO) V4. Specifically, first, the lightweight architecture MobileNetV3 is introduced as the backbone feature extractor to accelerate the detection speed by compressing the parameters. Second, for better detection of small-size ships, the squeeze-and-excitation module is used to apply the attention mechanism to the channel. Meanwhile, the scaled exponential linear unit non-liner activation function replaces the rectified linear unit activation function of the MobileNetV3 shallow layer, which optimizes the convergence effect of the model. Third, an adaptive anchor-selection algorithm for the detection of ships with various shapes is designed. Compared with other well-established models based on convolutional neural network (CNN), including single shot multi-box detector, Faster region based-CNN, and you only look once version 4 baseline for detecting ships, our improved method yields impressive results in our dataset. After extensive testing, the mean average precision of the proposed method can reach 97.43%, with the detection time per frame reaching 38 ms.

Analytical study of two feature extraction methods in comparison with deep learning methods for classification of small metal objects

This paper addresses the efficiency of two feature extraction methods for classifying small metal objects including screws, nuts, keys, and coins: the histogram of oriented gradients (HOG) and local binary pattern (LBP). The desired features for the labeled images are first extracted and saved in the form of a feature matrix. Using three different classification methods (non-parametric K-nearest neighbors algorithm, support vector machine, and naïve Bayesian method), the images are classified into four different classes. Then, by examining the resulting confusion matrix, the performances of the HOG and LBP approaches are compared for these four classes. The effectiveness of these two methods is also compared with the “You Only Look Once” and faster region-based convolutional neural network approaches, which are based on deep learning. The collected image set in this paper includes 800 labeled training images and 180 test images. The results show that the use of the HOG is more efficient than the use of the LBP. Moreover, a combination of the HOG and LBP provides better results than either alone.

Measuring inter-individual differences in behavioural types of gilthead seabreams in the laboratory using deep learning.

Deep learning allows us to automatize the acquisition of large amounts of behavioural animal data with applications for fisheries and aquaculture. In this work, we have trained an image-based deep learning algorithm, the Faster R-CNN (Faster region-based convolutional neural network), to automatically detect and track the gilthead seabream, Sparus aurata, to search for individual differences in behaviour. We collected videos using a novel Raspberry Pi high throughput recording system attached to individual experimental behavioural arenas. From the continuous recording during behavioural assays, we acquired and labelled a total of 14,000 images and used them, along with data augmentation techniques, to train the network. Then, we evaluated the performance of our network at different training levels, increasing the number of images and applying data augmentation. For every validation step, we processed more than 52,000 images, with and without the presence of the gilthead seabream, in normal and altered (i.e., after the introduction of a non-familiar object to test for explorative behaviour) behavioural arenas. The final and best version of the neural network, trained with all the images and with data augmentation, reached an accuracy of 92,79% ± 6.78% [89.24–96.34] of correct classification and 10.25 ± 61.59 pixels [6.59-13.91] of fish positioning error. Our recording system based on a Raspberry Pi and a trained convolutional neural network provides a valuable non-invasive tool to automatically track fish movements in experimental arenas and, using the trajectories obtained during behavioural tests, to assay behavioural types.

Automatic bleeding detection in laparoscopic surgery based on a faster region-based convolutional neural network.

BackgroundLaparoscopic surgery has been in great demand over the past decades; it has also brought several obstacles, such as increasing difficulty in maintaining hemostasis, changes in surgical approach, and reduced field of vision. Locating the bleeding point can help surgeons to control bleeding quickly, however, to date, there have been no tools designed for automatic bleeding tracking in laparoscopic operations. Herein, we have proposed a spatiotemporal hybrid model based on a faster region-based convolutional neural network (RCNN) for bleeding point detection in laparoscopic surgery videos.MethodsLaparoscopic videos performed at our hospital were retrieved and images containing bleeding events were extracted. Spatiotemporal features were extracted by using red-green-blue (RGB) frames and optical flow maps and a spatiotemporal hybrid model was developed based on the faster RCNN. The proposed model contributed to (I) providing real-time bleeding point detection which directly assist surgeons, (II) showing the blood’s optical flow which improved bleeding point detection, and (III) detecting both arterial and venous bleeding.ResultsIn this study, 12 different bleeding videos were included for deep learning model training. Compared with models containing a single RGB or a single optical flow map, our model combining RGB and optical flow achieved great detection results (precision rate of 0.8373, recall rate of 0.8034, and average precision of 0.6818).ConclusionsOur approach performs well in bleeding point location and recognition, indicating its potential value in helping to maintain and re-establish hemostasis during operations.

A novel deep learning method for marine oil spill detection from satellite synthetic aperture radar imagery

Oil spill discharges from operational maritime activities like ships, oil rigs and other structures, leaking pipelines, as well as natural hydrocarbon seepage pose serious threats to marine ecosystems and fisheries. Satellite synthetic aperture radar (SAR) is a unique microwave instrument for marine oil spill monitoring, as it is not dependent on weather or sunlight conditions. Existing SAR oil spill detection approaches are limited by algorithm complexity, imbalanced data sets, uncertainties in selecting optimal features, and relatively slow detection speed. To overcome these restrictions, a fast and effective SAR oil spill detection method is presented, based a novel deep learning model, named the Faster Region-based Convolutional Neural Network (Faster R-CNN). This approach is capable of achieving fast end-to-end oil spill detection with reasonable accuracy. A large data set consisting of 15,774 labeled oil spill samples derived from 1786C-band Sentinel-1 and RADARSAT-2 vertical polarization SAR images is used to train, validate and test the Faster R-CNN model. Our experimental results show that the proposed method exhibits good performance for detection of oil spills with wide swath SAR imagery. The Precision and Recall metrics are 89.23% and 89.14%, respectively. The average Precision is 92.56%. The effects of environmental conditions and sensor parameters on oil spill detection are analyzed. The expected detection results are obtained when wind speeds and incidence angles are between 3 m/s and 10 m/s, and 21° and 45°, respectively. Furthermore, the computer runtime for oil spill detection is less than 0.05 s for each full SAR image, using a workstation with NVIDIA GeForce RTX 3090 GPU. This suggests that the present approach has potential for applications that require fast oil spill detection from spaceborne SAR images.

Endoscopy Artefact Detection by Deep Transfer Learning of Baseline Models.

To visualise the tumours inside the body on a screen, a long and thin tube is inserted with a light source and a camera at the tip to obtain video frames inside organs in endoscopy. However, multiple artefacts exist in these video frames that cause difficulty during the diagnosis of cancers. In this research, deep learning was applied to detect eight kinds of artefacts: specularity, bubbles, saturation, contrast, blood, instrument, blur, and imaging artefacts. Based on transfer learning with pre-trained parameters and fine-tuning, two state-of-the-art methods were applied for detection: faster region-based convolutional neural networks (Faster R-CNN) and EfficientDet. Experiments were implemented on the grand challenge dataset, Endoscopy Artefact Detection and Segmentation (EAD2020). To validate our approach in this study, we used phase I of 2,200 frames and phase II of 331 frames in the original training dataset with ground-truth annotations as training and testing dataset, respectively. Among the tested methods, EfficientDet-D2 achieves a score of 0.2008 (mAPd[Formula: see text]0.6+mIoUd[Formula: see text]0.4) on the dataset that is better than three other baselines: Faster-RCNN, YOLOv3, and RetinaNet, and competitive to the best non-baseline result scored 0.25123 on the leaderboard although our testing was on phase II of 331 frames instead of the original 200 testing frames. Without extra improvement techniques beyond basic neural networks such as test-time augmentation, we showed that a simple baseline could achieve state-of-the-art performance in detecting artefacts in endoscopy. In conclusion, we proposed the combination of EfficientDet-D2 with suitable data augmentation and pre-trained parameters during fine-tuning training to detect the artefacts in endoscopy.

Weed Identification by Single-Stage and Two-Stage Neural Networks: A Study on the Impact of Image Resizers and Weights Optimization Algorithms.

The accurate identification of weeds is an essential step for a site-specific weed management system. In recent years, deep learning (DL) has got rapid advancements to perform complex agricultural tasks. The previous studies emphasized the evaluation of advanced training techniques or modifying the well-known DL models to improve the overall accuracy. In contrast, this research attempted to improve the mean average precision (mAP) for the detection and classification of eight classes of weeds by proposing a novel DL-based methodology. First, a comprehensive analysis of single-stage and two-stage neural networks including Single-shot MultiBox Detector (SSD), You look only Once (YOLO-v4), EfficientDet, CenterNet, RetinaNet, Faster Region-based Convolutional Neural Network (RCNN), and Region-based Fully Convolutional Network (RFCN), has been performed. Next, the effects of image resizing techniques along with four image interpolation methods have been studied. It led to the final stage of the research through optimization of the weights of the best-acquired model by initialization techniques, batch normalization, and DL optimization algorithms. The effectiveness of the proposed work is proven due to a high mAP of 93.44% and validated by the stratified k-fold cross-validation technique. It was 5.8% improved as compared to the results obtained by the default settings of the best-suited DL architecture (Faster RCNN ResNet-101). The presented pipeline would be a baseline study for the research community to explore several tasks such as real-time detection and reducing the computation/training time. All the relevant data including the annotated dataset, configuration files, and inference graph of the final model are provided with this article. Furthermore, the selection of the DeepWeeds dataset shows the robustness/practicality of the study because it contains images collected in a real/complex agricultural environment. Therefore, this research would be a considerable step toward an efficient and automatic weed control system.

A video-based SlowFastMTB model for detection of small amounts of smoke from incipient forest fires

Abstract This paper proposes a video-based SlowFast model that combines the SlowFast deep learning model with a new boundary box annotation algorithm. The new algorithm, namely the MTB (i.e., the ratio of the number of Moving object pixels To the number of Bounding box pixels) algorithm, is devised to automatically annotate the bounding box that includes the smoke with fuzzy boundaries. The model parameters of the MTB algorithm are examined by multifactor analysis of variance. To demonstrate the validity of the proposed approach, a case study is provided that examines real video clips of incipient forest fires with small amounts of smoke. The performance of the proposed approach is compared with those of existing deep learning models, including convolutional neural network (CNN), faster region-based CNN (faster R-CNN), and SlowFast. It is demonstrated that the proposed approach achieves enhanced detection accuracy, while reducing false negative rates.

Model for deep learning-based skill transfer in an assembly process

As the variety of products and manufacturing processes increases, the expansion of flexible training approaches is crucial to support the development of human skills. This study presents a model for skill transfer support that extracts experts’ relevant skills as actions and objects relevant to the action into a computational model for transferring skills. This model engages two modes of deep learning as the groundwork, namely, convolutional neural network (CNN) for action recognition and faster region-based convolutional neural network (R-CNN) for object detection. To evaluate the performance of the proposed model, a case study of the final assembly of a GPU card is conducted. The accuracy of CNN and faster R-CNN are 95.4% and 96.8%, respectively. The goal of this model is to guide junior operators during the assembly by providing step-by-step instructions in performing complex tasks. The present study facilitates flexible training in terms of adapting new skills from skilled operators to naïve operators by deep learning.

Automated prediction of the Thoracolumbar Injury Classification and Severity Score from CT using a novel deep learning algorithm

Damage to the thoracolumbar spine can confer significant morbidity and mortality. The Thoracolumbar Injury Classification and Severity Score (TLICS) is used to categorize injuries and determine patients at risk of spinal instability for whom surgical intervention is warranted. However, calculating this score can constitute a bottleneck in triaging and treating patients, as it relies on multiple imaging studies and a neurological examination. Therefore, the authors sought to develop and validate a deep learning model that can automatically categorize vertebral morphology and determine posterior ligamentous complex (PLC) integrity, two critical features of TLICS, using only CT scans. All patients who underwent neurosurgical consultation for traumatic spine injury or degenerative pathology resulting in spine injury at a single tertiary center from January 2018 to December 2019 were retrospectively evaluated for inclusion. The morphology of injury and integrity of the PLC were categorized on CT scans. A state-of-the-art object detection region-based convolutional neural network (R-CNN), Faster R-CNN, was leveraged to predict both vertebral locations and the corresponding TLICS. The network was trained with patient CT scans, manually labeled vertebral bounding boxes, TLICS morphology, and PLC annotations, thus allowing the model to output the location of vertebrae, categorize their morphology, and determine the status of PLC integrity. A total of 111 patients were included (mean ± SD age 62 ± 20 years) with a total of 129 separate injury classifications. Vertebral localization and PLC integrity classification achieved Dice scores of 0.92 and 0.88, respectively. Binary classification between noninjured and injured morphological scores demonstrated 95.1% accuracy. TLICS morphology accuracy, the true positive rate, and positive injury mismatch classification rate were 86.3%, 76.2%, and 22.7%, respectively. Classification accuracy between no injury and suspected PLC injury was 86.8%, while true positive, false negative, and false positive rates were 90.0%, 10.0%, and 21.8%, respectively. In this study, the authors demonstrate a novel deep learning method to automatically predict injury morphology and PLC disruption with high accuracy. This model may streamline and improve diagnostic decision support for patients with thoracolumbar spinal trauma.

Cervical Cancer Diagnostics Healthcare System Using Hybrid Object Detection Adversarial Networks

Cervical cancer is one of the common cancers among women and it causes significant mortality in many developing countries. Diagnosis of cervical lesions is done using pap smear test or visual inspection using acetic acid (staining). Digital colposcopy, an inexpensive methodology, provides painless and efficient screening results. Therefore, automating cervical cancer screening using colposcopy images will be highly useful in saving many lives. Nowadays, many automation techniques using computer vision and machine learning in cervical screening gained attention, paving the way for diagnosing cervical cancer. However, most of the methods rely entirely on the annotation of cervical spotting and segmentation. This paper aims to introduce the Faster Small-Object Detection Neural Networks (FSOD-GAN) to address the cervical screening and diagnosis of cervical cancer and the type of cancer using digital colposcopy images. The proposed approach automatically detects the cervical spot using Faster Region-Based Convolutional Neural Network (FR-CNN) and performs the hierarchical multiclass classification of three types of cervical cancer lesions. Experimentation was done with colposcopy data collected from available open sources consisting of 1,993 patients with three cervical categories, and the proposed approach shows 99% accuracy in diagnosing the stages of cervical cancer.

High-throughput measurement of plant fitness traits with an object detection method using Faster R-CNN.

Summary Revealing the contributions of genes to plant phenotype is frequently challenging because loss‐of‐function effects may be subtle or masked by varying degrees of genetic redundancy. Such effects can potentially be detected by measuring plant fitness, which reflects the cumulative effects of genetic changes over the lifetime of a plant. However, fitness is challenging to measure accurately, particularly in species with high fecundity and relatively small propagule sizes such as Arabidopsis thaliana.An image segmentation‐based method using the software ImageJ and an object detection‐based method using the Faster Region‐based Convolutional Neural Network (R‐CNN) algorithm were used for measuring two Arabidopsis fitness traits: seed and fruit counts.The segmentation‐based method was error‐prone (correlation between true and predicted seed counts, r 2 = 0.849) because seeds touching each other were undercounted. By contrast, the object detection‐based algorithm yielded near perfect seed counts (r 2 = 0.9996) and highly accurate fruit counts (r 2 = 0.980). Comparing seed counts for wild‐type and 12 mutant lines revealed fitness effects for three genes; fruit counts revealed the same effects for two genes.Our study provides analysis pipelines and models to facilitate the investigation of Arabidopsis fitness traits and demonstrates the importance of examining fitness traits when studying gene functions.

Faster CNN-based vehicle detection and counting strategy for fixed camera scenes

Automatic detection and counting of vehicles in a video is a challenging task and has become a key application area of traffic monitoring and management. In this paper, an efficient real-time approach for the detection and counting of moving vehicles is presented based on YOLOv2 and features point motion analysis. The work is based on synchronous vehicle features detection and tracking to achieve accurate counting results. The proposed strategy works in two phases; the first one is vehicle detection and the second is the counting of moving vehicles. Different convolutional neural networks including pixel by pixel classification networks and regression networks are investigated to improve the detection and counting decisions. For initial object detection, we have utilized state-of-the-art faster deep learning object detection algorithm YOLOv2 before refining them using K-means clustering and KLT tracker. Then an efficient approach is introduced using temporal information of the detection and tracking feature points between the framesets to assign each vehicle label with their corresponding trajectories and truly counted it. Experimental results on twelve challenging videos have shown that the proposed scheme generally outperforms state-of-the-art strategies. Moreover, the proposed approach using YOLOv2 increases the average time performance for the twelve tested sequences by 93.4% and 98.9% from 1.24 frames per second achieved using Faster Region-based Convolutional Neural Network (F R-CNN ) and 0.19 frames per second achieved using the background subtraction based CNN approach (BS-CNN ), respectively to 18.7 frames per second.

Evaluation of a deep learning system for automatic detection of proximal surface dental caries on bitewing radiographs

This study aimed to evaluate a deep learning (DL) system using convolutional neural networks (CNNs) for automatic detection of caries on bitewing radiographs. In total, 2468 bitewings were labeled by 3 dentists to create the reference standard. Of these images, 1257 had caries and 1211 were sound. The Faster region-based CNN was applied to detect the regions of interest (ROIs) with potential lesions. A total of 13,246 ROIs were generated from all 'sound' images, and 50% of 'caries' images (selected randomly) were used to train the ROI detection module. The remaining 50% of 'caries' images were used to validate the ROI detection module. Caries detection was then performed using Inception-ResNet-v2. A set of 3297 'caries' and 5321 'sound' ROIs cropped from the 2468 images was used to train and validate the caries detection module. Data sets were randomly divided into training (90%) and validation (10%) data sets. Recall, precision, specificity, accuracy, and F1 score were used as metrics to assess performance. The caries detection module achieved recall, precision, specificity, accuracy, and F1 scores of 0.89, 0.86, 0.86, 0.87, and 0.87, respectively. The proposed DL system demonstrated promising performance for detecting proximal surface caries on bitewings.