Predicting osteoporosis from kidney-ureter-bladder radiographs utilizing deep convolutional neural networks.

Osteoporosis is a common condition that can lead to fractures, mobility issues, and death. Although dual-energy X-ray absorptiometry (DXA) is the gold standard for osteoporosis, it is expensive and not widely available. In contrast, kidney-ureter-bladder (KUB) radiographs are inexpensive and frequently ordered in clinical practice. Thus, it is a potential screening tool for osteoporosis. In this study, we explored the possibility of predicting the bone mineral density (BMD) and classifying high-risk patient groups using KUB radiographs. We proposed DeepDXA-KUB, a deep learning model that predicts the BMD values of the left hip and lumbar vertebrae from an input KUB image. The datasets were obtained from Taiwanese medical centers between 2006 and 2019, using 8913 pairs of KUB radiographs and DXA examinations performed within 6 months. The images were randomly divided into training and validation sets in a 4:1 ratio. To evaluate the model's performance, we computed a confusion matrix and evaluated the sensitivity, specificity, accuracy, precision, positive predictive value, negative predictive value, F1 score, and area under the receiver operating curve (AUROC). Moderate correlations were observed between the predicted and DXA-measured BMD values, with a correlation coefficient of 0.858 for the lumbar vertebrae and 0.87 for the left hip. The model demonstrated an osteoporosis detection accuracy, sensitivity, and specificity of 84.7 %, 81.6 %, and 86.6 % for the lumbar vertebrae and 84.2 %, 91.2 %, and 81 % for the left hip, respectively. The AUROC was 0.939 for the lumbar vertebrae and 0.947 for the left hip, indicating a satisfactory performance in osteoporosis screening. The present study is the first to develop a deep learning model based on KUB radiographs to predict lumbar spine and femoral BMD. Our model demonstrated a promising correlation between the predicted and DXA-measured BMD in both the lumbar vertebrae and hip, showing great potential for the opportunistic screening of osteoporosis.

Diagnostic Accuracy of Deep Learning for the Prediction of Osteoporosis Using Plain X-rays: A Systematic Review and Meta-Analysis.

(1) Background: This meta-analysis assessed the diagnostic accuracy of deep learning model-based osteoporosis prediction using plain X-ray images. (2) Methods: We searched PubMed, Web of Science, SCOPUS, and Google Scholar from no set beginning date to 28 February 2023, for eligible studies that applied deep learning methods for diagnosing osteoporosis using X-ray images. The quality of studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 criteria. The area under the receiver operating characteristic curve (AUROC) was used to quantify the predictive performance. Subgroup, meta-regression, and sensitivity analyses were performed to identify the potential sources of study heterogeneity. (3) Results: Six studies were included; the pooled AUROC, sensitivity, and specificity were 0.88 (95% confidence interval [CI] 0.85-0.91), 0.81 (95% CI 0.78-0.84), and 0.87 (95% CI 0.81-0.92), respectively, indicating good performance. Moderate heterogeneity was observed. Mega-regression and subgroup analyses were not performed due to the limited number of studies included. (4) Conclusion: Deep learning methods effectively extract bone density information from plain radiographs, highlighting their potential for opportunistic screening. Nevertheless, additional prospective multicenter studies involving diverse patient populations are required to confirm the applicability of this novel technique.

Automatic and human level Graf's type identification for detecting developmental dysplasia of the hip.

BACKGROUND: Developmental dysplasia of the hip (DDH) is a common congenital disorder that may lead to hip dislocation and requires surgical intervention if left untreated. Ultrasonography is the preferred method for DDH screening; however, the lack of experienced operators impedes its application in universal neonatal screening. METHODS: We developed a deep neural network tool to automatically register the five keypoints that mark important anatomical structures of the hip and provide a reference for measuring alpha and beta angles following Graf's guidelines, which is an ultrasound classification system for DDH in infants. Two-dimensional (2D) ultrasonography images were obtained from 986 neonates aged 0-6 months. A total of 2406 images from 921 patients were labeled with ground truth keypoints by senior orthopedists. RESULTS: Our model demonstrated precise keypoint localization. The mean absolute error was approximately 1 mm, and the derived alpha angle measurement had a correlation coefficient of R = 0.89 between the model and ground truth. The model achieved an area under the receiver operating characteristic curve of 0.937 and 0.974 for classifying alpha <60° (abnormal hip) and <50° (dysplastic hip), respectively. On average, the experts agreed with 96% of the inferenced images, and the model could generalize its prediction on newly collected images with a correlation coefficient higher than 0.85. CONCLUSIONS: Precise localization and highly correlated performance metrics suggest that the model can be an efficient tool for assisting DDH diagnosis in clinical settings.

Recovery of walking ability in stroke patients through postacute care rehabilitation.

BACKGROUND: Walking entails orchestration of the sensory, motor, balance, and coordination systems, and walking disability is a critical concern after stroke. How and to what extent these systems influence walking disability after stroke and recovery have not been comprehensively studied. METHODS: We retrospectively analyzed patients with stroke in the Post-acute care-Cerebrovascular Diseases (PAC-CVD) program. We compared the characteristics of patient groups stratified by their ability to complete the 5-m walk test across various time points of rehabilitation. We then used stepwise linear regression to examine the degree to which each stroke characteristic and functional ability could predict patient gait performance. RESULTS: Five hundred seventy-three patients were recruited, and their recovery of walking ability was defined by the timing of recovery in a 5-m walk test. The proportion of patients who could complete the 5-m walk test at admission, at 3 weeks of rehabilitation, at 6 weeks of rehabilitation, between 7 and 12 weeks of rehabilitation, and who could not complete the 5-m walk test after rehabilitation was 52.2%, 21.8%, 8.7%, 8.7%, and 8.6%, respectively. At postacute care discharge, patients who regained walking ability earlier had a higher chance of achieving higher levels of walking activity. Stepwise linear regression showed that Berg Balance Scale (BBS) (ß: 0.011, p < .001), age (ß: -0.005, p = .001), National Institutes of Health Stroke Scale (NIHSS) (6a + 6b; ß: -0.042, p = .018), Mini-Nutritional assessment (MNA) (ß: -0.007, p < .027), and Fugl-Meyer upper extremity assessment (FuglUE) (ß: 0.002, p = .047) scores predicted patient's gait speed at discharge. CONCLUSION: Balance, age, leg strength, nutritional status, and upper limb function before postacute care rehabilitation are predictors of walking performance after stroke.

Efficacy of machine learning assisted dental age assessment in local population.

INTRODUCTION: Although the dental age assessment is commonly applied in forensic and maturity evaluation, the long-standing dilemma from population differences has limited its application. OBJECTIVES: This study aimed to verify the efficacy of the machine learning (ML) to build up the dental age standard of a local population. METHODS: We retrospectively studied 2052 panoramic films retrieved from healthy Taiwanese children aged 2.6-17.7 years with comparable sizes in each age-group. The recently reported Han population-based standard (H method) served as the control condition. To develop and validate ML models, random divisions of the sample in an 80%-20% ratio repeated 20 times. The model performances were compared with the H method, Demirjian's method, and Willems's method. RESULTS: The ML-assisted models provided more accurate age prediction than those non-ML-assisted methods. The range of errors was effectively reduced to less than one per year in the ML models. Furthermore, the consistent agreements among the age groups from preschool to adolescence were reported for the first time. The Gaussian process regression was the best ML model; of the non-ML modalities, the H method was the most efficacious, followed by the Demirjian's method and Willems's methods. CONCLUSION: The ML-assisted dental age assessment is helpful to provide customized standards to a local population with more accurate estimations in preschool and adolescent age groups than do studied conventional methods. In addition, the earlier complete tooth developments were also observed in present study. To construct more reliable dental maturity models in the future, additional environment-related factors should be taken into account.

The Application of Artificial-Intelligence-Assisted Dental Age Assessment in Children with Growth Delay.

BACKGROUND: This study aimed to reveal the efficacy of the artificial intelligence (AI)-assisted dental age (DA) assessment in identifying the characteristics of growth delay (GD) in children. METHODS: The panoramic films matching the inclusion criteria were collected for the AI model training to establish the population-based DA standard. Subsequently, the DA of the validation dataset of the healthy children and the images of the GD children were assessed by both the conventional methods and the AI-assisted standards. The efficacy of all the studied modalities was compared by the paired sample t-test. RESULTS: The AI-assisted standards can provide much more accurate chronological age (CA) predictions with mean errors of less than 0.05 years, while the traditional methods presented overestimated results in both genders. For the GD children, the convolutional neural network (CNN) revealed the delayed DA in GD children of both genders, while the machine learning models presented so only in the GD boys. CONCLUSION: The AI-assisted DA assessments help overcome the long-standing populational limitation observed in traditional methods. The image feature extraction of the CNN models provided the best efficacy to reveal the nature of delayed DA in GD children of both genders.

Deep Learning-Based Grimace Scoring Is Comparable to Human Scoring in a Mouse Migraine Model.

Pain assessment is essential for preclinical and clinical studies on pain. The mouse grimace scale (MGS), consisting of five grimace action units, is a reliable measurement of spontaneous pain in mice. However, MGS scoring is labor-intensive and time-consuming. Deep learning can be applied for the automatic assessment of spontaneous pain. We developed a deep learning model, the DeepMGS, that automatically crops mouse face images, predicts action unit scores and total scores on the MGS, and finally infers whether pain exists. We then compared the performance of DeepMGS with that of experienced and apprentice human scorers. The DeepMGS achieved an accuracy of 70-90% in identifying the five action units of the MGS, and its performance (correlation coefficient = 0.83) highly correlated with that of an experienced human scorer in total MGS scores. In classifying pain and no pain conditions, the DeepMGS is comparable to the experienced human scorer and superior to the apprentice human scorers. Heatmaps generated by gradient-weighted class activation mapping indicate that the DeepMGS accurately focuses on MGS-relevant areas in mouse face images. These findings support that the DeepMGS can be applied for quantifying spontaneous pain in mice, implying its potential application for predicting other painful conditions from facial images.

Automatic prosthetic-parameter estimation from anteroposterior pelvic radiographs after total hip arthroplasty using deep learning-based keypoint detection.

BACKGROUND: X-ray is a necessary tool for post-total hip arthroplasty (THA) check-ups; however, parameter measurements are time-consuming. We proposed a deep learning tool, BKNet that automates localization of landmarks with parameter measurements. METHODS: About 3072 radiographs from 3021 patients who underwent THA at our institute between 2013 and 2017 were used. We employed BKNet to perform landmark localization with parameter measurements in these radiographs. The performance of BKNet was assessed and compared with that of human observers. RESULTS: The 75-percentile cut-off errors were <0.5 cm in all key points. The Bland-Altman methods show the agreement between the predicted and ground truth parameters. Human and BKNet comparison revealed the model could match the repeatability for 7/10 of the parameters. CONCLUSIONS: The accuracy of BKNet is equivalent to that of human observers, and BKNet was able to perform prosthetic-parameter estimation from keypoint detection with superior cost-effectiveness, repeatability, and timesaving compared to human observers.

The Comparison and Interpretation of Machine-Learning Models in Post-Stroke Functional Outcome Prediction.

Prediction of post-stroke functional outcomes is crucial for allocating medical resources. In this study, a total of 577 patients were enrolled in the Post-Acute Care-Cerebrovascular Disease (PAC-CVD) program, and 77 predictors were collected at admission. The outcome was whether a patient could achieve a Barthel Index (BI) score of >60 upon discharge. Eight machine-learning (ML) methods were applied, and their results were integrated by stacking method. The area under the curve (AUC) of the eight ML models ranged from 0.83 to 0.887, with random forest, stacking, logistic regression, and support vector machine demonstrating superior performance. The feature importance analysis indicated that the initial Berg Balance Test (BBS-I), initial BI (BI-I), and initial Concise Chinese Aphasia Test (CCAT-I) were the top three predictors of BI scores at discharge. The partial dependence plot (PDP) and individual conditional expectation (ICE) plot indicated that the predictors' ability to predict outcomes was the most pronounced within a specific value range (e.g., BBS-I < 40 and BI-I < 60). BI at discharge could be predicted by information collected at admission with the aid of various ML models, and the PDP and ICE plots indicated that the predictors could predict outcomes at a certain value range.

Application of deep learning neural network in predicting bone mineral density from plain X-ray radiography.

DeepDXA is a deep learning model designed to infer bone mineral density data from plain pelvis X-ray, and it can achieve good predicted value for clinical use. PURPOSE: Osteoporosis is defined as a systemic disease of the bone characterized by a decrease in bone strength and deterioration of bone structure at the microscopic level, leading to bone fragility and increased risk of fracture. Bone mineral density (BMD) is the preferred method for the diagnosis of osteoporosis, and dual-energy x-ray absorptiometry (DXA) is the gold standard for diagnosing osteoporosis. Conventional radiography is more suited for the screening of osteoporosis rather than diagnosis, and osteoporosis can be detected on radiographs by experienced physicians only. This study explored the possibility of predicting BMD relative to DXA using patient radiographs. METHODS: A deep learning algorithm of convolutional neural network (CNN) was used for the purpose. The method includes image segmentation, CNN learning, and a convolution-based regression model (DeepDXA) that links the isolated images of the femur bone to predict BMD value. Data were obtained in a single medical center from 2006 to 2018, with a total amount of 3472 pairs of pelvis X-ray and DXA examination within 1 year. RESULTS: The proposed workflow successfully predicted BMD values of the femur bone with the correlation coefficient (R) of 0.85 (P < 0.001) and the accuracy of 0.88 for prediction osteoporosis, a finding that could be reliably ready for further clinical use. CONCLUSION: When suspicious osteoporosis is seen on plain films using the deep learning method we developed, further referral to DXA for the definite diagnosis of osteoporosis is indicated.

A Multimodal Imaging-Based Deep Learning Model for Detecting Treatment-Requiring Retinal Vascular Diseases: Model Development and Validation Study.

BACKGROUND: Retinal vascular diseases, including diabetic macular edema (DME), neovascular age-related macular degeneration (nAMD), myopic choroidal neovascularization (mCNV), and branch and central retinal vein occlusion (BRVO/CRVO), are considered vision-threatening eye diseases. However, accurate diagnosis depends on multimodal imaging and the expertise of retinal ophthalmologists. OBJECTIVE: The aim of this study was to develop a deep learning model to detect treatment-requiring retinal vascular diseases using multimodal imaging. METHODS: This retrospective study enrolled participants with multimodal ophthalmic imaging data from 3 hospitals in Taiwan from 2013 to 2019. Eye-related images were used, including those obtained through retinal fundus photography, optical coherence tomography (OCT), and fluorescein angiography with or without indocyanine green angiography (FA/ICGA). A deep learning model was constructed for detecting DME, nAMD, mCNV, BRVO, and CRVO and identifying treatment-requiring diseases. Model performance was evaluated and is presented as the area under the curve (AUC) for each receiver operating characteristic curve. RESULTS: A total of 2992 eyes of 2185 patients were studied, with 239, 1209, 1008, 211, 189, and 136 eyes in the control, DME, nAMD, mCNV, BRVO, and CRVO groups, respectively. Among them, 1898 eyes required treatment. The eyes were divided into training, validation, and testing groups in a 5:1:1 ratio. In total, 5117 retinal fundus photos, 9316 OCT images, and 20,922 FA/ICGA images were used. The AUCs for detecting mCNV, DME, nAMD, BRVO, and CRVO were 0.996, 0.995, 0.990, 0.959, and 0.988, respectively. The AUC for detecting treatment-requiring diseases was 0.969. From the heat maps, we observed that the model could identify retinal vascular diseases. CONCLUSIONS: Our study developed a deep learning model to detect retinal diseases using multimodal ophthalmic imaging. Furthermore, the model demonstrated good performance in detecting treatment-requiring retinal diseases.

Functional Recovery Patterns of Hemorrhagic and Ischemic Stroke Patients Under Post-Acute Care Rehabilitation Program.

BACKGROUND: Tailored rehabilitation programs for stroke patients cannot be made without knowledge of their recovery potential. The aim of this study is to characterize the functional recovery patterns of ischemic stroke (IS) and intracerebral hemorrhage (ICH) patients under post-acute care stroke rehabilitation. METHODS: This retrospective study analyzed the data of patients enrolled in the Post-Acute Care-Cerebrovascular Disease (PAC-CVD) rehabilitation program, which provides an individualized 1- to 3-hour intensive physical, occupational, and speech and language therapy for post-acute stroke patients in Taoyuan Chang Gung Memorial hospital in Taiwan. Our primary endpoint measure was Barthel Index (BI), and secondary endpoint measures included other 12 functional measures. RESULTS: A total of 489 patients were included for analysis. Patients with stroke history had less BI improvement than those who suffered their first-ever stroke. In first-ever stroke patients who had completed 6 to 12 weeks of PAC-CVD program, subcortical ICH patients had greater BI, quality of life, sensation, and balance improvements, and had greater late-phase recovery than their IS counterparts. In IS patients, those with age >75 had less BI improvement; those with National Institute of Health Stroke Scale (NIHSS) score 1-5 had greater Motor Activity Log quality of use (MAL-quality) improvement than those with NIHSS score >5; those with Mini-Mental State Examination (MMSE) score ≥24 had greater BI and instrumental activities of daily living (IADL) improvement. Using the general linear model, previous stroke (ß: -6.148, p=0.01) and subcortical ICH (ß: 5.04, p=0.03) were factors associated with BI improvement. CONCLUSION: Subcortical ICH patients have greater functional improvement and greater late-phase recovery than their IS counterparts following PAC rehabilitation. More studies are needed to validate our findings and unravel the underlying mechanisms of stroke recovery to optimize the treatment strategy following a stroke.

Author Correction: Relative posture between head and finger determines perceived tactile direction of motion.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

Relative posture between head and finger determines perceived tactile direction of motion.

The hand explores the environment for obtaining tactile information that can be fruitfully integrated with other functions, such as vision, audition, and movement. In theory, somatosensory signals gathered by the hand are accurately mapped in the world-centered (allocentric) reference frame such that the multi-modal information signals, whether visual-tactile or motor-tactile, are perfectly aligned. However, an accumulating body of evidence indicates that the perceived tactile orientation or direction is inaccurate; yielding a surprisingly large perceptual bias. To investigate such perceptual bias, this study presented tactile motion stimuli to healthy adult participants in a variety of finger and head postures, and requested the participants to report the perceived direction of motion mapped on a video screen placed on the frontoparallel plane in front of the eyes. Experimental results showed that the perceptual bias could be divided into systematic and nonsystematic biases. Systematic bias, defined as the mean difference between the perceived and veridical directions, correlated linearly with the relative posture between the finger and the head. By contrast, nonsystematic bias, defined as minor difference in bias for different stimulus directions, was highly individualized, phase-locked to stimulus orientation presented on the skin. Overall, the present findings on systematic bias indicate that the transformation bias among the reference frames is dominated by the finger-to-head posture. Moreover, the highly individualized nature of nonsystematic bias reflects how information is obtained by the orientation-selective units in the S1 cortex.

Network Anisotropy Trumps Noise for Efficient Object Coding in Macaque Inferior Temporal Cortex.

How neuronal ensembles compute information is actively studied in early visual cortex. Much less is known about how local ensembles function in inferior temporal (IT) cortex, the last stage of the ventral visual pathway that supports visual recognition. Previous reports suggested that nearby neurons carry information mostly independently, supporting efficient processing (Barlow, 1961). However, others postulate that noise covariation effects may depend on network anisotropy/homogeneity and on how the covariation relates to representation. Do slow trial-by-trial noise covariations increase or decrease IT's object coding capability, how does encoding capability relate to correlational structure (i.e., the spatial pattern of signal and noise redundancy/homogeneity across neurons), and does knowledge of correlational structure matter for decoding? We recorded simultaneously from ∼80 spiking neurons in ∼1 mm(3) of macaque IT under light neurolept anesthesia. Noise correlations were stronger for neurons with correlated tuning, and noise covariations reduced object encoding capability, including generalization across object pose and illumination. Knowledge of noise covariations did not lead to better decoding performance. However, knowledge of anisotropy/homogeneity improved encoding and decoding efficiency by reducing the number of neurons needed to reach a given performance level. Such correlated neurons were found mostly in supragranular and infragranular layers, supporting theories that link recurrent circuitry to manifold representation. These results suggest that redundancy benefits manifold learning of complex high-dimensional information and that subsets of neurons may be more immune to noise covariation than others. SIGNIFICANCE STATEMENT: How noise affects neuronal population coding is poorly understood. By sampling densely from local populations supporting visual object recognition, we show that recurrent circuitry supports useful representations and that subsets of neurons may be more immune to noise covariation than others.

Corrigendum: Correlated activity supports efficient cortical processing.

[This corrects the article on p. 171 in vol. 8, PMID: 25610392.].

Tuning and spontaneous spike time synchrony share a common structure in macaque inferior temporal cortex.

Investigating the relationship between tuning and spike timing is necessary to understand how neuronal populations in anterior visual cortex process complex stimuli. Are tuning and spontaneous spike time synchrony linked by a common spatial structure (do some cells covary more strongly, even in the absence of visual stimulation?), and what is the object coding capability of this structure? Here, we recorded from spiking populations in macaque inferior temporal (IT) cortex under neurolept anesthesia. We report that, although most nearby IT neurons are weakly correlated, neurons with more similar tuning are also more synchronized during spontaneous activity. This link between tuning and synchrony was not simply due to cell separation distance. Instead, it expands on previous reports that neurons along an IT penetration are tuned to similar but slightly different features. This constraint on possible population firing rate patterns was consistent across stimulus sets, including animate vs. inanimate object categories. A classifier trained on this structure was able to generalize category "read-out" to untrained objects using only a few dimensions (a few patterns of site weightings per electrode array). We suggest that tuning and spike synchrony are linked by a common spatial structure that is highly efficient for object representation.

Correlated activity supports efficient cortical processing.

Visual recognition is a computational challenge that is thought to occur via efficient coding. An important concept is sparseness, a measure of coding efficiency. The prevailing view is that sparseness supports efficiency by minimizing redundancy and correlations in spiking populations. Yet, we recently reported that "choristers", neurons that behave more similarly (have correlated stimulus preferences and spontaneous coincident spiking), carry more generalizable object information than uncorrelated neurons ("soloists") in macaque inferior temporal (IT) cortex. The rarity of choristers (as low as 6% of IT neurons) indicates that they were likely missed in previous studies. Here, we report that correlation strength is distinct from sparseness (choristers are not simply broadly tuned neurons), that choristers are located in non-granular output layers, and that correlated activity predicts human visual search efficiency. These counterintuitive results suggest that a redundant correlational structure supports efficient processing and behavior.