小橋 昌司

Maximizing healthy life expectancy is essential for enhancing well-being. Optimal exercise intensity is crucial in promoting health and ensuring safe rehabilitation. Since heart rate is related to exercise intensity, the required exercise intensity is achieved by controlling the heart rate. This study aims to control heart rate during exercise by dynamically adjusting the load on a bicycle ergometer using a proportional-integral (PI) control. The choice of PI parameters is very important because the PI parameters significantly affect the performance of heart rate control. Since the dynamic characteristics of heart rate relative to work rate vary widely from subject to subject, the PI parameters for each subject must be determined individually. In this study, PI parameters are optimized directly from exercise data using a data-driven design approach. Thus, the proposed method does not require excessive exercise of the subject to model heart rate dynamics. Using the proposed method, the heart rate can be controlled to follow a designed reference model so that the heart rate is safely increased to the desired value. The quantitative evaluation of the control results of fifteen healthy volunteers confirmed that the proposed method improved the control error of the target heart rate trajectory by approximately 40%, regardless of gender or age. In addition, it was shown that control parameters from the exercise experiment also indicate that females are more likely than males to have an elevated heart rate at the same load.

BACKGROUND: Ultrasonography is used to diagnose osteochondritis dissecans (OCD) of the humerus; however, its reliability depends on the technical proficiency of the examiner. Recently, computer-aided diagnosis (CAD) using deep learning has been applied in the field of medical science, and high diagnostic accuracy has been reported. We aimed to develop a deep learning-based CAD system for OCD detection on ultrasound images and to evaluate the accuracy of OCD detection using the CAD system. METHODS: The CAD process comprises 2 steps: humeral capitellum detection using an object-detection algorithm and OCD classification using an image classification network. Four-directional ultrasound images of the elbow of the throwing arm of 196 baseball players (mean age, 11.2 years), including 104 players with normal findings and 92 with OCD, were used for training and validation. An external dataset of 20 baseball players (10 with normal findings and 10 with OCD) was used to evaluate the accuracy of the CAD system. A confusion matrix and the area under the receiver operating characteristic curve (AUC) were used to evaluate the system. RESULTS: Clinical evaluation using the external dataset resulted in high AUCs in all 4 directions: 0.969 for the anterior long axis, 0.966 for the anterior short axis, 0.996 for the posterior long axis, and 0.993 for the posterior short axis. The accuracy of OCD detection thus exceeded 0.9 in all 4 directions. CONCLUSIONS: We propose a deep learning-based CAD system to detect OCD lesions on ultrasound images. The CAD system achieved high accuracy in all 4 directions of the elbow. This CAD system with a deep learning model may be useful for OCD screening during medical checkups to reduce the probability of missing an OCD lesion. LEVEL OF EVIDENCE: Diagnostic Level II. See Instructions for Authors for a complete description of levels of evidence.

Pelvic fractures pose significant challenges in medical diagnosis due to the complex structure of the pelvic bones. Timely diagnosis of pelvic fractures is critical to reduce complications and mortality rates. While computed tomography (CT) is highly accurate in detecting pelvic fractures, the initial diagnostic procedure usually involves pelvic X-rays (PXR). In recent years, many deep learning-based methods have been developed utilizing ImageNet-based transfer learning for diagnosing hip and pelvic fractures. However, the ImageNet dataset contains natural RGB images which are different than PXR. In this study, we proposed a two-step transfer learning approach that improved the diagnosis of pelvic fractures in PXR images. The first step involved training a deep convolutional neural network (DCNN) using synthesized PXR images derived from 3D-CT by digitally reconstructed radiographs (DRR). In the second step, the classification layers of the DCNN were fine-tuned using acquired PXR images. The performance of the proposed method was compared with the conventional ImageNet-based transfer learning method. Experimental results demonstrated that the proposed DRR-based method, using 20 synthesized PXR images for each CT, achieved superior performance with the area under the receiver operating characteristic curves (AUROCs) of 0.9327 and 0.8014 for visible and invisible fractures, respectively. The ImageNet-based method yields AUROCs of 0.8908 and 0.7308 for visible and invisible fractures, respectively.

PURPOSE: Osteochondritis dissecans (OCD) of the humeral capitellum is a common cause of elbow disorders, particularly among young throwing athletes. Conservative treatment is the preferred treatment for managing OCD, and early intervention significantly influences the possibility of complete disease resolution. The purpose of this study is to develop a deep learning-based classification model in ultrasound images for computer-aided diagnosis. METHODS: This paper proposes a deep learning-based OCD classification method in ultrasound images. The proposed method first detects the humeral capitellum detection using YOLO and then estimates the OCD probability of the detected region probability using VGG16. We hypothesis that the performance will be improved by eliminating unnecessary regions. To validate the performance of the proposed method, it was applied to 158 subjects (OCD: 67, Normal: 91) using five-fold-cross-validation. RESULTS: The study demonstrated that the humeral capitellum detection achieved a mean average precision (mAP) of over 0.95, while OCD probability estimation achieved an average accuracy of 0.890, precision of 0.888, recall of 0.927, F1 score of 0.894, and an area under the curve (AUC) of 0.962. On the other hand, when the classification model was constructed for the entire image, accuracy, precision, recall, F1 score, and AUC were 0.806, 0.806, 0.932, 0.843, and 0.928, respectively. The findings suggest the high-performance potential of the proposed model for OCD classification in ultrasonic images. CONCLUSION: This paper introduces a deep learning-based OCD classification method. The experimental results emphasize the effectiveness of focusing on the humeral capitellum for OCD classification in ultrasound images. Future work should involve evaluating the effectiveness of employing the proposed method by physicians during medical check-ups for OCD.

Pelvic and hip fractures offer considerable public health risks with high morbidity and mortality rates. Because of the complicated bone structure of the pelvic bone region, detecting fractures is difficult. Though X-ray imaging is routinely utilised for detecting fractures, manual fracture diagnosis is prone to inaccuracies. This paper proposes the use of deep learning algorithms for automated segmentation of the pelvic bone region in X-ray images. In our work, we have investigated U-Net based pelvic area segmentation models with various convolutional neural network (CNN) backbones. The DenseNet121-based U-Net design emerged as the most optimal model, establishing a compromise between performance and computational efficiency. Although it had a modest loss in IoU and F1 scores when compared to InceptionNetV3, it had a remarkable 59.44% reduction in the number of parameters.

Abstract Deep learning techniques for automatically detecting teeth in dental X-rays have gained popularity, providing valuable assistance to healthcare professionals. However, teeth detection in X-ray images is often hindered by alterations in tooth appearance caused by dental prostheses. To address this challenge, our paper proposes a novel method for teeth detection and numbering in dental panoramic X-rays, leveraging two separate CNN-based object detectors, namely YOLOv7, for detecting teeth and prostheses, alongside an optimization algorithm to refine the outcomes. The study utilizes a dataset of 3138 radiographs, of which 2553 images contain prostheses, to build a robust model. The tooth and prosthesis detection algorithms perform excellently, achieving mean average precisions of 0.982 and 0.983, respectively. Additionally, the trained tooth detection model is verified using an external dataset, and six-fold cross-validation is conducted to demonstrate the proposed method’s feasibility and robustness. Moreover, the investigation of performance improvement resulting from the inclusion of prosthesis information in the teeth detection process reveals a marginal increase in the average F1-score, rising from 0.985 to 0.987 compared to the sole teeth detection method. The proposed method is unique in its approach to numbering teeth as it incorporates prosthesis information and considers complete restorations such as dental implants and dentures of fixed bridges during the teeth enumeration process, which follows the universal tooth numbering system. These advancements hold promise for automating dental charting processes.

The incidence of falls in hospital facilities is high and can lead to a decrease in patients' quality of life and an increase in medical expenses. Therefore, the development of a system that can predict getting-up from a bed is necessary. This study proposes a get-up detecting sensor using 6-axis inertial sensor. This system can detect getting-up from a bed in real-time using machine learning with Edge AI. To evaluate the basic performance of the proposed system, a protocol was applied for four subjects, and data were collected. Alert accuracy rate and false alert rate were used as evaluation metrics, and a model was built using data from three of the subjects and evaluated with the remaining subject, which was repeated for all four subjects. For high-risk bed-leaving behavior, medium-risk pre-bed-leaving behavior, and low-risk get-up behavior, the alert accuracy rate (i.e., Recall) was 89.4%, 97.5%, and 86.3%, respectively, and the false alert rate (1-Precision) was 6.3%, 10.2%, and 0.0%, respectively. This confirmed the possibility of predicting rising behavior with high accuracy. Furthermore, as a proof of concept, a real-time get-up detection system was developed, and its practicality was demonstrated. Future challenges include reevaluating feature extraction and evaluating the performance of the proposed system with a diverse range of subjects of different ages, genders, and health statuses.

Kidney tumor is a health concern that affects kidney cells and may leads to mortality depending on their type. Benign tumors can be unproblematic whereas malignant tumors pose the threat of kidney cancer. Early detection and diagnosis are possible through kidney tumor recognition based on deep learning techniques. In this paper, a method based on transfer learning using deep convolutional neural network (DCNN) is proposed to recognize kidney tumor from computed tomography (CT) images. The proposed method was evaluated on 5284 images. The final accuracy, precision, recall, specificity and F1 score were 92.54%, 80.45%, 93.02%, 92.38% and 0.8628, respectively.

The number of heart disease cases as well as the death associated with it are rising in numbers every year. It is now more important than ever to diagnose heart abnormalities quickly and correctly to ensure proper treatment is provided in time. A common tool for diagnosing heart abnormalities is the Electrocardiogram (ECG). The ECG is a procedure that requires electrodes to monitor and records the activity of hearts as a form of signal. In this paper, a method is proposed to classify standard 12-lead ECG signals using continuous wavelet transform (CWT) and convolutional neural network (CNN). At first, CWT is used to extract and represent features of the ECG signals in 2-dimensional (2D) RGB images. Later, the RGB images are classified into normal and abnormal cases using a pre-trained CNN. The proposed method is evaluated using a dataset containing ECG signals from 18,885 subjects. The maximum accuracy, precision, recall, F1 score, and AUC obtained are 74.78%, 78.968%, 71.003%, 72.957%, and 0.81126 respectively.

肩腱板断裂は，日常生活の動作やスポーツ外傷，加齢などによって引き起こされる一般的な肩関節の障害である．肩腱板断裂の診断には磁気共鳴画像装置（MRI）が広く用いられているが，より迅速かつ普及した撮像手法として単純X線撮影がある．本研究ではより簡便な肩腱板断裂診断のために，肩のX線画像に畳み込みニューラルネットワークを用いた検出法を提案した．その結果139名の被験者，断裂の有無による2クラス分類において，最大79.3%の検出精度を示した．

Diabetic sensorimotor polyneuropathy (DSPN) is a serious long-term complication of diabetes, which may lead to foot ulceration and amputation. Among the screening tools for DSPN, the Michigan neuropathy screening instrument (MNSI) is frequently deployed, but it lacks a straightforward rating of severity. A DSPN severity grading system has been built and simulated for the MNSI, utilizing longitudinal data captured over 19 years from the Epidemiology of Diabetes Interventions and Complications (EDIC) trial. Machine learning algorithms were used to establish the MNSI factors and patient outcomes to characterise the features with the best ability to detect DSPN severity. A nomogram based on multivariable logistic regression was designed, developed and validated. The extra tree model was applied to identify the top seven ranked MNSI features that identified DSPN, namely vibration perception (R), 10-gm filament, previous diabetic neuropathy, vibration perception (L), presence of callus, deformities and fissure. The nomogram’s area under the curve (AUC) was 0.9421 and 0.946 for the internal and external datasets, respectively. The probability of DSPN was predicted from the nomogram and a DSPN severity grading system for MNSI was created using the probability score. An independent dataset was used to validate the model’s performance. The patients were divided into four different severity levels, i.e., absent, mild, moderate, and severe, with cut-off values of 10.50, 12.70 and 15.00 for a DSPN probability of less than 50, 75 and 100%, respectively. We provide an easy-to-use, straightforward and reproducible approach to determine prognosis in patients with DSPN.

Intracerebral hematoma (ICH) is a blood clot that forms when a blood vessel in the brain ruptures for some reason and the spilled blood coagulates. ICH has a high morbidity and mortality rate, accounting for approximately 10% of all strokes. Manual segmentation of ICH in head CT images is very complicated, time consuming, and troublesome. When ICH perforates the ventricular wall and blood flows into the ventricle, there is little difference in CT value between ICH and intraventricular hemorrhage (IVH), and the boundary between them is unclear. Convolutional neural network (CNN) has proven to be a reliable method in the field of image recognition. In addition, quantification of ICH may aid in decision making in ICH treatment. In this study, we introduce CNN in a stepwise manner to differentiate ICH and IVH in the process and extract ICH regions. The results in 18 stroke patients show that our method is promising in the extraction of ICH regions with an accuracy of 75.2%.

The purpose of this study is to analyze clinical features and radiomics features obtained from MR images of the carotid arteries to evaluate the accuracy of predicting whether asymptomatic plaque would migrate to symptomatic plaque. Using the extracted radiomics features and clinical features, and machine learning algorithms are used to predict symptomatic migration. The machine learning algorithms used are SVM, Logistic regression, LightGBM, and Random Forest. Their performances are evaluated. Prediction by clinical features alone was AUC 0.709, and a combination of clinical and radiomics features was AUC 0.744 respectively. LightGBM showed the best accuracy. The combined features model showed effectiveness in predicting carotid symptomatology migration.

In recent years, healthcare and safety have been a major focus of deep learning research. This paper focuses on the detection of Medical Personal Protective Equipment (MPPE) in the health-care sector using YOLOv7. Improper use of personal protective equipment (PPE) can result in the contamination and cross-contamination of infectious diseases, so it is crucial for healthcare professionals to use it correctly. The CPPE-5 dataset was used to train the model, which contains 1029 high-quality images divided into five categories: coveralls, face shield, gloves, masks, and goggles. The objective of this research is to create an accurate model for future applications and development using a suitable medical PPE dataset. The proposed model outperforms previous studies, with an optimal mAP of 90.93 %, indicating that it is a promising method for detecting MPPE in the healthcare sector.

Fragility fracture of pelvis (FFP) is increasingly affecting elderly population. Although computed tomography (CT) imaging is considered superior to conventional radiographic image for diagnosing FFP, clinicians face challenges in recognizing pelvic fractures owing to imaging contrast or feature size. This study proposes a method that combines boring survey based FFP candidate extraction from CT images and a newly developed convolutional neural network model. In addition, the proposed method also visualizes the probability of fracture on 3D bone surface data. The accuracy, precision, and recall of the proposed method were found to be 79.7%, 60.0%, and 80.6%, respectively. Furthermore, the 3D view of fracture probability on the pelvic bone surface allows for qualitative assessment and can support physicians to diagnose FFP. The findings indicate that the proposed method has potential for predicting FFP.

Neonatal chronic lung disease (CLD) is the most common and serious lung disease in premature infants. No previous studies have used chest X-ray images. In this study, we propose to predict and classify patients with and without CLD from neonatal chest X-ray images using a convolutional neural network (CNN). We conducted a 5-segment cross-validation experiment using chest X-ray images of 115 subjects at 7 days of age. Accuracy and AUC values of 0.6 and 0.642 were obtained, respectively. Future work includes the development of an algorithm suitable for neonatal data and the estimation of other age groups.

Baseball elbow is a kinetic disorder of the elbow caused by repetitive pitching in baseball. Osteochondritis dissecans (OCD), one of the most common forms of baseball elbow, is a disorder of the humeral capitellum of the elbow, and early detection of OCD is important. This study aims to create a model to detect OCD from ultrasound images of the elbow. The model is based on VGG16. The proposed method was validated by using 67 OCD subjects and 91 normal subjects. The results showed that the model achieved an accuracy of 88.5%, a precision of 87.9%, a recall of 97.0%, an F1 score of 0.910, and an AUC of 0.971.

Self-removal problems, in which patients pull out their own medical tubing, remain unresolved while medical tubing introduces more than before. Not only patients can't receive proper treatment by removing of medical tubing, but also complications such as bleeding and injury from the removal site are considered problematic. It has been adopted as a solution that frequent visits by nurses and monitoring patients, but their burden has become a problem. To reduce this issue, prior studies have examined systems for detecting medical tubing tension, but false positives were caused by patients turning over in bed or being cared for by nurses. Here we show that a prototype flexible sensor that independently detects medical tubing removal and tension. In the evaluation, the sensor detects tension by being pulled, and then applied a certain amount of force, it comes off. In addition, in order to monitor the status of medical tubing in real time, we implemented server transmission of sensor values via BLE (Bluetooth Low Energy), graphing, and LINE notification functions. Our results demonstrate that highly accurate medical tubing condition sensing is possible and reduces the burden on nurses and patients.

Brain development in children between the ages of 0 and 3 is extremely rapid. In the diagnosis of the pediatric brain, quantitative methods for evaluating brain growth during this period are needed. We propose a method for predicting the developmental age of the brain using deep learning. In the proposed method, at first, the cranial region is extracted from CT images, and then the posture and position are calibrated. We propose a new neural network model that uses a 3D convolutional neural network (3D-CNN) to extract features from CT images and estimate the developmental age of the brain in all coupled layers. 204 pediatric patients (0 to 47 months) with no neurological abnormalities were used for study and evaluation. The root mean square error (RMSE) between the predicted age and the patient's actual age was 6.45 months with a correlation coefficient of 0.89. In addition, the output of the attention map showed a high degree of attention to the anterior region. This result is consistent with medical findings that anterior regions of the pediatric brain are particularly developed.

Brain imaging is used to diagnose pediatric brain diseases. However, there is no quantitative method to estimate developmental conditions such as underdevelopment or early growth, and qualitative diagnosis is based on the experience of skilled physicians. Therefore, we are developing a computer-aided diagnosis system to estimate brain age from pediatric brain CT images. This system segmented cranial regions from CT images and calibrated their posture and position. The system also extracts features from CT images using a 3D convolutional neural network (3D CNN) and predicts brain age using a fully connected layer. This paper focuses on the cranial region segmentation method, which is an essential analysis processing method for the system. We investigated two different methods of region segmentation, and a comparison experiment with 204 subjects aged 0 to 3 years (47 months) showed that we could improve 32% of the prediction accuracy of the 3D CNN model.

Chronic lung disease (CLD) is the most common and serious lung disease in premature infants. In this study, we predict the severity (mild or severe) of neonatal chest X-ray images using a convolutional neural network (CNN) to enable early intervention to provide personalized treatment and improve prognosis. Thirty subjects were tested in a leave-one-out cross validation experiment using 30 chest X-ray images of 11 patients with mild disease and 19 patients with severe disease at 7 days of age. To improve the prediction accuracy, we proposed to limit the input image of the CNN to the lung field region and to use a pre-training model for transfer learning. Four different experiments were conducted, comparing the results with different input images (whole image or lung field region) and with and without transfer learning. The results showed that the best accuracy was obtained when the entire image was used as input and no transfer learning was performed, with an Accuracy of 0.667.

Baseball elbow is a pitching elbow disorder caused by repeated pitching movements. Osteochondritis dissecans (OCD) is one of baseball elbow disorders, and is an intractable osteochondral injury that tends to occur in elementary and junior high school students. If it can be found in the early stages, it will be completely cured by conservative treatment, which is to set a period to stop playing baseball. Since there is almost no pain in the early stage, the hurdles for consultation are high and there are many cases in which the condition becomes severe. Periodical medical check of baseball elbow is effective, however, the number of implementations is several times a year due to the shortage of specialists who can make a diagnosis. In this study, for the purpose of developing computer-aided diagnosis (CADx) of early-stage OCD, we propose an OCD detection method using ultrasound images of the elbow. The proposed method first segments the humerus capitellum using fully convolutional network (FCN). Secondly, the segmented region is classified into OCD +/- classes using fine-tuning VGG16 to detect OCD. The proposed method was applied to 125 child baseball players including 61 OCD children and 64 healthy children. 5-fold cross-validation was conducted. The average detection results were 76.8% for accuracy, 100% for precision, 52.3% for recall, F1-score was 0.673, and AUC was 0.851.

In the current dental practice, many panoramic dental images of the oral cavity are taken by x-ray radiograph. Using the dental panoramic images, a physician or dental assistant records dental chart. These burdens can deteriorate the quality of medical care, such as erroneous entries. Therefore, automatic analysis of panoramic dental images is desired. We have previously proposed a teeth recognition method based on Faster R-CNN and an optimization approach that performed a 94.2% accuracy. However, it shows a relatively low accuracy in panoramic images with prostheses. This paper proposed a new method to improve the accuracy by detecting prostheses separately. It first detects four types of prosthetic teeth using YOLOv5. Then, it recognizes the teeth and the prosthetic teeth simultaneously based on the proposed optimization approach using a prior knowledge model. The proposed method achieved a maximum recognition accuracy of 97.17%. It shows the usefulness of optimization using prior knowledge models in combination with prosthetic tooth detection.

Fragility fractures of the pelvis (FFP) often occur because of the osteoporosis in the elderly, and are difficult to found on transverse CT images. It desires an automatic pelvic fracture detection system for improving quality of medicine. Our previous study, which is based on 3D analysis using CT images of the whole pelvis, achieved a precision-recall area-under-the-curve (PR-AUC) of 0.824. However, the recall of impact fractures is relatively low (0.350). To overcome this difficulty, we have previously introduced a novel approach of complementary detecting pelvic impact fractures, called boring survey based fracture detection (BSFD). In order to improve the performance of detecting impact fractures based on BSFD, this study proposes a new 3D CNN model with considering the 3D feature architecture. The detection accuracy is improved by eliminating the separated fractures from the pelvic surface. The separated fractures are detected by our previous method. Also, with respect to the imbalance problem, the fractured regions are over-sampled in training data. To evaluate the proposed method, this study recruited 28 patients as the training data and 4 patients as the validation data. It achieved a receiver operating characteristic AUC (ROC-AUC) of 0.878 and a recall of 0.806 for detecting impact fractures for the validation data. We can also three-dimensionally visualize the estimated degree of fractures for a sake of easy understanding by physicians.

Approximately 600,000 to 1,000,000 patients are diagnosed with rheumatoid arthritis (RA) in Japan. To provide appropriate treatment, it is necessary to accurately measure the progression of RA by diagnosing the disease several times a year. The modified total sharp score (mTSS) calculated from hand X-ray images is a standard diagnostic method for RA progression. However, this diagnostic method is time-consuming as the scores are rated at as many as 16 points per hand. Accordingly, in order to shorten the diagnosis time of RA patients and improve the quality of diagnosis, the development of computer-aided diagnosis (CAD) systems is expected. We have previously proposed a CAD system that can detect finger joint positions using a support vector machine and can estimate the mTSS using ridge regression. In this study, we propose a fully automatic detection method of RA score evaluation points in the carpal site from simple hand X-ray images using deep learning. The proposed method first segments the carpal site using deep learning. Next, the RA evaluation points are automatically determined from each segment based on prior knowledge. Experimental results on X-ray images of the hands of 140 patients with RA showed that the mTSS evaluation point at the carpal site could be detected with an average error of 25 pixels. This study enables the automatic detection of RA score evaluation points in the carpal site. In the diagnosis of RA, the time required for diagnosis can be reduced by automating the determination of diagnostic points by physician.

BACKGROUND: Diabetic Sensorimotor polyneuropathy (DSPN) is one of the major indelible complications in diabetic patients. Michigan neuropathy screening instrumentation (MNSI) is one of the most common screening techniques used for DSPN, however, it does not provide any direct severity grading system. METHOD: For designing and modeling the DSPN severity grading systems for MNSI, 19 years of data from Epidemiology of Diabetes Interventions and Complications (EDIC) clinical trials were used. Different Machine learning-based feature ranking techniques were investigated to identify the important MNSI features associated with DSPN diagnosis. A multivariable logistic regression-based nomogram was generated and validated for DSPN severity grading using the best performing top-ranked MNSI features. RESULTS: Top-10 ranked features from MNSI features: Appearance of Feet (R), Ankle Reflexes (R), Vibration perception (L), Vibration perception (R), Appearance of Feet (L), 10-gm filament (L), Ankle Reflexes (L), 10-gm filament (R), Bed Cover Touch, and Ulceration (R) were identified as important features for identifying DSPN by Multi-Tree Extreme Gradient Boost model. The nomogram-based prediction model exhibited an accuracy of 97.95% and 98.84% for the EDIC test set and an independent test set, respectively. A DSPN severity score technique was generated for MNSI from the DSPN severity prediction model. DSPN patients were stratified into four severity levels: absent, mild, moderate, and severe using the cut-off values of 17.6, 19.1, 20.5 for the DSPN probability less than 50%, 75%-90%, and above 90%, respectively. CONCLUSIONS: The findings of this work provide a machine learning-based MNSI severity grading system which has the potential to be used as a secondary decision support system by health professionals in clinical applications and large clinical trials to identify high-risk DSPN patients.

Pelvic fracture is one of the leading causes of death in the elderly, carrying a high risk of death within 1 year of fracture. This study proposes an automated method to detect pelvic fractures on 3-dimensional computed tomography (3D-CT). Deep convolutional neural networks (DCNNs) have been used for lesion detection on 2D and 3D medical images. However, training a DCNN directly using 3D images is complicated, computationally costly, and requires large amounts of training data. We propose a method that evaluates multiple, 2D, real-time object detection systems (YOLOv3 models) in parallel, in which each YOLOv3 model is trained using differently orientated 2D slab images reconstructed from 3D-CT. We assume that an appropriate reconstruction orientation would exist to optimally characterize image features of bone fractures on 3D-CT. Multiple YOLOv3 models in parallel detect 2D fracture candidates in different orientations simultaneously. The 3D fracture region is then obtained by integrating the 2D fracture candidates. The proposed method was validated in 93 subjects with bone fractures. Area under the curve (AUC) was 0.824, with 0.805 recall and 0.907 precision. The AUC with a single orientation was 0.652. This method was then applied to 112 subjects without bone fractures to evaluate over-detection. The proposed method successfully detected no bone fractures in all except 4 non-fracture subjects (96.4%).