Critical evaluation of deep neural networks for wrist fracture detection

Diagnostic Performance of Artificial Intelligence for Detection of Scaphoid and Distal Radius Fractures: A Systematic Review

PURPOSE

To review the existing literature to (1) determine the diagnostic efficacy of artificial intelligence (AI) models for detecting scaphoid and distal radius fractures and (2) compare the efficacy to human clinical experts.

METHODS

PubMed, OVID/Medline, and Cochrane libraries were queried for studies investigating the development, validation, and analysis of AI for the detection of scaphoid or distal radius fractures. Data regarding study design, AI model development and architecture, prediction accuracy/area under the receiver operator characteristic curve (AUROC), and imaging modalities were recorded.

RESULTS

A total of 21 studies were identified, of which 12 (57.1%) used AI to detect fractures of the distal radius, and nine (42.9%) used AI to detect fractures of the scaphoid. AI models demonstrated good diagnostic performance on average, with AUROC values ranging from 0.77 to 0.96 for scaphoid fractures and from 0.90 to 0.99 for distal radius fractures. Accuracy of AI models ranged between 72.0% to 90.3% and 89.0% to 98.0% for scaphoid and distal radius fractures, respectively. When compared to clinical experts, 13 of 14 (92.9%) studies reported that AI models demonstrated comparable or better performance. The type of fracture influenced model performance, with worse overall performance on occult scaphoid fractures; however, models trained specifically on occult fractures demonstrated substantially improved performance when compared to humans.

CONCLUSIONS

AI models demonstrated excellent performance for detecting scaphoid and distal radius fractures, with the majority demonstrating comparable or better performance compared with human experts. Worse performance was demonstrated on occult fractures. However, when trained specifically on difficult fracture patterns, AI models demonstrated improved performance.

CLINICAL RELEVANCE

AI models can help detect commonly missed occult fractures while enhancing workflow efficiency for distal radius and scaphoid fracture diagnoses. As performance varies based on fracture type, future studies focused on wrist fracture detection should clearly define whether the goal is to (1) identify difficult-to-detect fractures or (2) improve workflow efficiency by assisting in routine tasks.

Deep learning performance compared to healthcare experts in detecting wrist fractures from radiographs: A systematic review and meta-analysis

OBJECTIVE

To perform a systematic review and meta-analysis of the diagnostic accuracy of deep learning (DL) algorithms in the diagnosis of wrist fractures (WF) on plain wrist radiographs, taking healthcare experts consensus as reference standard.

METHODS

Embase, Medline, PubMed, Scopus and Web of Science were searched in the period from 1 Jan 2012 to 9 March 2023. Eligible studies were patients with wrist radiographs for radial and ulnar fractures as the target condition, studies using DL algorithms based on convolutional neural networks (CNN), and healthcare experts consensus as the minimum reference standard. Studies were assessed with a modified QUADAS-2 tool, and we applied a bivariate random-effects model for meta-analysis of diagnostic test accuracy data.

RESULTS

Our study was registered at PROSPERO with ID: CRD42023431398. We included 6 unique studies for meta-analysis, with a total of 33,026 radiographs. CNN performance compared to reference standards for the included articles found a summary sensitivity of 92% (95% CI: 80%-97%) and a summary specificity of 93% (95% CI: 76%-98%). The generalized bivariate I-squared statistic indicated considerable heterogeneity between the studies (81.90%). Four studies had one or more domains at high risk of bias and two studies had concerns regarding applicability.

CONCLUSION

The diagnostic accuracy of CNNs was comparable to that of healthcare experts in wrist radiographs for investigation of WF. There is a need for studies with a robust reference standard, external data-set validation and investigation of diagnostic performance of healthcare experts aided with CNNs.

CLINICAL RELEVANCE STATEMENT

DL matches healthcare experts in diagnosing WFs, which potentially benefits patient diagnosis.

Diagnostic power of ChatGPT 4 in distal radius fracture detection through wrist radiographs

Distal radius fractures rank among the most prevalent fractures in humans, necessitating accurate radiological imaging and interpretation for optimal diagnosis and treatment. In addition to human radiologists, artificial intelligence systems are increasingly employed for radiological assessments. Since 2023, ChatGPT 4 has offered image analysis capabilities, which can also be used for the analysis of wrist radiographs. This study evaluates the diagnostic power of ChatGPT 4 in identifying distal radius fractures, comparing it with a board-certified radiologist, a hand surgery resident, a medical student, and the well-established AI Gleamer BoneView™. Results demonstrate ChatGPT 4's good diagnostic accuracy (sensitivity 0.88, specificity 0.98, diagnostic power (AUC) 0.93), surpassing the medical student (sensitivity 0.98, specificity 0.72, diagnostic power (AUC) 0.85; p = 0.04) significantly. Nevertheless, the diagnostic power of ChatGPT 4 lags behind the hand surgery resident (sensitivity 0.99, specificity 0.98, diagnostic power (AUC) 0.985; p = 0.014) and Gleamer BoneView™(sensitivity 1.00, specificity 0.98, diagnostic power (AUC) 0.99; p = 0.006). This study highlights the utility and potential applications of artificial intelligence in modern medicine, emphasizing ChatGPT 4 as a valuable tool for enhancing diagnostic capabilities in the field of medical imaging.

Artificial intelligence in fracture detection with different image modalities and data types: A systematic review and meta-analysis

Artificial Intelligence (AI), encompassing Machine Learning and Deep Learning, has increasingly been applied to fracture detection using diverse imaging modalities and data types. This systematic review and meta-analysis aimed to assess the efficacy of AI in detecting fractures through various imaging modalities and data types (image, tabular, or both) and to synthesize the existing evidence related to AI-based fracture detection. Peer-reviewed studies developing and validating AI for fracture detection were identified through searches in multiple electronic databases without time limitations. A hierarchical meta-analysis model was used to calculate pooled sensitivity and specificity. A diagnostic accuracy quality assessment was performed to evaluate bias and applicability. Of the 66 eligible studies, 54 identified fractures using imaging-related data, nine using tabular data, and three using both. Vertebral fractures were the most common outcome (n = 20), followed by hip fractures (n = 18). Hip fractures exhibited the highest pooled sensitivity (92%; 95% CI: 87-96, p< 0.01) and specificity (90%; 95% CI: 85-93, p< 0.01). Pooled sensitivity and specificity using image data (92%; 95% CI: 90-94, p< 0.01; and 91%; 95% CI: 88-93, p < 0.01) were higher than those using tabular data (81%; 95% CI: 77-85, p< 0.01; and 83%; 95% CI: 76-88, p < 0.01), respectively. Radiographs demonstrated the highest pooled sensitivity (94%; 95% CI: 90-96, p < 0.01) and specificity (92%; 95% CI: 89-94, p< 0.01). Patient selection and reference standards were major concerns in assessing diagnostic accuracy for bias and applicability. AI displays high diagnostic accuracy for various fracture outcomes, indicating potential utility in healthcare systems for fracture diagnosis. However, enhanced transparency in reporting and adherence to standardized guidelines are necessary to improve the clinical applicability of AI. Review Registration: PROSPERO (CRD42021240359).

Development and optimization of AI algorithms for wrist fracture detection in children using a freely available dataset

Introduction

In the field of pediatric trauma computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems have emerged offering a promising avenue for improved patient care. Especially children with wrist fractures may benefit from machine learning (ML) solutions, since some of these lesions may be overlooked on conventional X-ray due to minimal compression without dislocation or mistaken for cartilaginous growth plates. In this article, we describe the development and optimization of AI algorithms for wrist fracture detection in children.

Methods

A team of IT-specialists, pediatric radiologists and pediatric surgeons used the freely available GRAZPEDWRI-DX dataset containing annotated pediatric trauma wrist radiographs of 6,091 patients, a total number of 10,643 studies (20,327 images). First, a basic object detection model, a You Only Look Once object detector of the seventh generation (YOLOv7) was trained and tested on these data. Then, team decisions were taken to adjust data preparation, image sizes used for training and testing, and configuration of the detection model. Furthermore, we investigated each of these models using an Explainable Artificial Intelligence (XAI) method called Gradient Class Activation Mapping (Grad-CAM). This method visualizes where a model directs its attention to before classifying and regressing a certain class through saliency maps.

Results

Mean average precision (mAP) improved when applying optimizations pre-processing the dataset images (maximum increases of + 25.51% mAP@0.5 and + 39.78% mAP@[0.5:0.95]), as well as the object detection model itself (maximum increases of + 13.36% mAP@0.5 and + 27.01% mAP@[0.5:0.95]). Generally, when analyzing the resulting models using XAI methods, higher scoring model variations in terms of mAP paid more attention to broader regions of the image, prioritizing detection accuracy over precision compared to the less accurate models.

Discussion

This paper supports the implementation of ML solutions for pediatric trauma care. Optimization of a large X-ray dataset and the YOLOv7 model improve the model's ability to detect objects and provide valid diagnostic support to health care specialists. Such optimization protocols must be understood and advocated, before comparing ML performances against health care specialists.

Deep Learning Model Based on You Only Look Once Algorithm for Detection and Visualization of Fracture Areas in Three-Dimensional Skeletal Images

Utilizing "You only look once" (YOLO) v4 AI offers valuable support in fracture detection and diagnostic decision-making. The purpose of this study was to help doctors to detect and diagnose fractures more accurately and intuitively, with fewer errors. The data accepted into the backbone are diversified through CSPDarkNet-53. Feature maps are extracted using Spatial Pyramid Pooling and a Path Aggregation Network in the neck part. The head part aggregates and generates the final output. All bounding boxes by the YOLO v4 are mapped onto the 3D reconstructed bone images after being resized to match the same region as shown in the 2D CT images. The YOLO v4-based AI model was evaluated through precision-recall (PR) curves and the intersection over union (IoU). Our proposed system facilitated an intuitive display of the fractured area through a distinctive red mask overlaid on the 3D reconstructed bone images. The high average precision values (>0.60) were reported as 0.71 and 0.81 from the PR curves of the tibia and elbow, respectively. The IoU values were calculated as 0.6327 (tibia) and 0.6638 (elbow). When utilized by orthopedic surgeons in real clinical scenarios, this AI-powered 3D diagnosis support system could enable a quick and accurate trauma diagnosis.

Automated computation of radiographic parameters of distal radial metaphyseal fractures in forearm X-rays

PURPOSE

Radiographic parameters (RPs) provide objective support for effective decision making in determining clinical treatment of distal radius fractures (DRFs). This paper presents a novel automatic RP computation pipeline for computing the six anatomical RPs associated with DRFs in anteroposterior (AP) and lateral (LAT) forearm radiographs.

METHODS

The pipeline consists of: (1) segmentation of the distal radius and ulna bones with six 2D Dynamic U-Net deep learning models; (2) landmark points detection and distal radius axis computation from the segmentations with geometric methods; (3) RP computation and generation of a quantitative DRF report and composite AP and LAT radiograph images. This hybrid approach combines the advantages of deep learning and model-based methods.

RESULTS

The pipeline was evaluated on 90 AP and 93 LAT radiographs for which ground truth distal radius and ulna segmentations and RP landmarks were manually obtained by expert clinicians. It achieves an accuracy of 94 and 86% on the AP and LAT RPs, within the observer variability, and an RP measurement difference of 1.4 ± 1.2° for the radial angle, 0.5 ± 0.6 mm for the radial length, 0.9 ± 0.7 mm for the radial shift, 0.7 ± 0.5 mm for the ulnar variance, 2.9 ± 3.3° for the palmar tilt and 1.2 ± 1.0 mm for the dorsal shift.

CONCLUSION

Our pipeline is the first fully automatic method that accurately and robustly computes the RPs for a wide variety of clinical forearm radiographs from different sources, hand orientations, with and without cast. The computed accurate and reliable RF measurements may support fracture severity assessment and clinical management.

Assessing the Potential of a Deep Learning Tool to Improve Fracture Detection by Radiologists and Emergency Physicians on Extremity Radiographs

RATIONALE AND OBJECTIVES

To evaluate the standalone performance of a deep learning (DL) based fracture detection tool on extremity radiographs and assess the performance of radiologists and emergency physicians in identifying fractures of the extremities with and without the DL aid.

MATERIALS AND METHODS

The DL tool was previously developed using 132,000 appendicular skeletal radiographs divided into 87% training, 11% validation, and 2% test sets. Stand-alone performance was evaluated on 2626 de-identified radiographs from a single institution in Ohio, including at least 140 exams per body region. Consensus from three US board-certified musculoskeletal (MSK) radiologists served as ground truth. A multi-reader retrospective study was performed in which 24 readers (eight each of emergency physicians, non-MSK radiologists, and MSK radiologists) identified fractures in 186 cases during two independent sessions with and without DL aid, separated by a one-month washout period. The accuracy (area under the receiver operating curve), sensitivity, specificity, and reading time were compared with and without model aid.

RESULTS

The model achieved a stand-alone accuracy of 0.986, sensitivity of 0.987, and specificity of 0.885, and high accuracy (> 0.95) across stratification for body part, age, gender, radiographic views, and scanner type. With DL aid, reader accuracy increased by 0.047 (95% CI: 0.034, 0.061; p = 0.004) and sensitivity significantly improved from 0.865 (95% CI: 0.848, 0.881) to 0.955 (95% CI: 0.944, 0.964). Average reading time was shortened by 7.1 s (27%) per exam. When stratified by physician type, this improvement was greater for emergency physicians and non-MSK radiologists.

CONCLUSION

The DL tool demonstrated high stand-alone accuracy, aided physician diagnostic accuracy, and decreased interpretation time.

Commercially-available AI algorithm improves radiologists' sensitivity for wrist and hand fracture detection on X-ray, compared to a CT-based ground truth

OBJECTIVES

Algorithms for fracture detection are spreading in clinical practice, but the use of X-ray-only ground truth can induce bias in their evaluation. This study assessed radiologists' performances to detect wrist and hand fractures on radiographs, using a commercially-available algorithm, compared to a computerized tomography (CT) ground truth.

METHODS

Post-traumatic hand and wrist CT and concomitant X-ray examinations were retrospectively gathered. Radiographs were labeled based on CT findings. The dataset was composed of 296 consecutive cases: 118 normal (39.9%), 178 pathological (60.1%) with a total of 267 fractures visible in CT. Twenty-three radiologists with various levels of experience reviewed all radiographs without AI, then using it, blinded towards CT results.

RESULTS

Using AI improved radiologists' sensitivity (Se, 0.658 to 0.703, p < 0.0001) and negative predictive value (NPV, 0.585 to 0.618, p < 0.0001), without affecting their specificity (Sp, 0.885 vs 0.891, p = 0.91) or positive predictive value (PPV, 0.887 vs 0.899, p = 0.08). On the radiographic dataset, based on the CT ground truth, stand-alone AI performances were 0.771 (Se), 0.898 (Sp), 0.684 (NPV), 0.915 (PPV), and 0.764 (AUROC) which were lower than previously reported, suggesting a potential underestimation of the number of missed fractures in the AI literature.

CONCLUSIONS

AI enabled radiologists to improve their sensitivity and negative predictive value for wrist and hand fracture detection on radiographs, without affecting their specificity or positive predictive value, compared to a CT-based ground truth. Using CT as gold standard for X-ray labels is innovative, leading to algorithm performance poorer than reported elsewhere, but probably closer to clinical reality.

CLINICAL RELEVANCE STATEMENT

Using an AI algorithm significantly improved radiologists' sensitivity and negative predictive value in detecting wrist and hand fractures on radiographs, with ground truth labels based on CT findings.

KEY POINTS

• Using CT as a ground truth for labeling X-rays is new in AI literature, and led to algorithm performance significantly poorer than reported elsewhere (AUROC: 0.764), but probably closer to clinical reality. • AI enabled radiologists to significantly improve their sensitivity (+ 4.5%) and negative predictive value (+ 3.3%) for the detection of wrist and hand fractures on X-rays. • There was no significant change in terms of specificity or positive predictive value.

FracAtlas: A Dataset for Fracture Classification, Localization and Segmentation of Musculoskeletal Radiographs

Digital radiography is one of the most common and cost-effective standards for the diagnosis of bone fractures. For such diagnoses expert intervention is required which is time-consuming and demands rigorous training. With the recent growth of computer vision algorithms, there is a surge of interest in computer-aided diagnosis. The development of algorithms demands large datasets with proper annotations. Existing X-Ray datasets are either small or lack proper annotation, which hinders the development of machine-learning algorithms and evaluation of the relative performance of algorithms for classification, localization, and segmentation. We present FracAtlas, a new dataset of X-Ray scans curated from the images collected from 3 major hospitals in Bangladesh. Our dataset includes 4,083 images that have been manually annotated for bone fracture classification, localization, and segmentation with the help of 2 expert radiologists and an orthopedist using the open-source labeling platform, makesense.ai. There are 717 images with 922 instances of fractures. Each of the fracture instances has its own mask and bounding box, whereas the scans also have global labels for classification tasks. We believe the dataset will be a valuable resource for researchers interested in developing and evaluating machine learning algorithms for bone fracture diagnosis.

Comparison of diagnostic performance of a deep learning algorithm, emergency physicians, junior radiologists and senior radiologists in the detection of appendicular fractures in children

BACKGROUND

Advances have been made in the use of artificial intelligence (AI) in the field of diagnostic imaging, particularly in the detection of fractures on conventional radiographs. Studies looking at the detection of fractures in the pediatric population are few. The anatomical variations and evolution according to the child's age require specific studies of this population. Failure to diagnose fractures early in children may lead to serious consequences for growth.

OBJECTIVE

To evaluate the performance of an AI algorithm based on deep neural networks toward detecting traumatic appendicular fractures in a pediatric population. To compare sensitivity, specificity, positive predictive value and negative predictive value of different readers and the AI algorithm.

MATERIALS AND METHODS

This retrospective study conducted on 878 patients younger than 18 years of age evaluated conventional radiographs obtained after recent non-life-threatening trauma. All radiographs of the shoulder, arm, elbow, forearm, wrist, hand, leg, knee, ankle and foot were evaluated. The diagnostic performance of a consensus of radiology experts in pediatric imaging (reference standard) was compared with those of pediatric radiologists, emergency physicians, senior residents and junior residents. The predictions made by the AI algorithm and the annotations made by the different physicians were compared.

RESULTS

The algorithm predicted 174 fractures out of 182, corresponding to a sensitivity of 95.6%, a specificity of 91.64% and a negative predictive value of 98.76%. The AI predictions were close to that of pediatric radiologists (sensitivity 98.35%) and that of senior residents (95.05%) and were above those of emergency physicians (81.87%) and junior residents (90.1%). The algorithm identified 3 (1.6%) fractures not initially seen by pediatric radiologists.

CONCLUSION

This study suggests that deep learning algorithms can be useful in improving the detection of fractures in children.

Detection of incomplete atypical femoral fracture on anteroposterior radiographs via explainable artificial intelligence

One of the key aspects of the diagnosis and treatment of atypical femoral fractures is the early detection of incomplete fractures and the prevention of their progression to complete fractures. However, an incomplete atypical femoral fracture can be misdiagnosed as a normal lesion by both primary care physicians and orthopedic surgeons; expert consultation is needed for accurate diagnosis. To overcome this limitation, we developed a transfer learning-based ensemble model to detect and localize fractures. A total of 1050 radiographs, including 100 incomplete fractures, were preprocessed by applying a Sobel filter. Six models (EfficientNet B5, B6, B7, DenseNet 121, MobileNet V1, and V2) were selected for transfer learning. We then composed two ensemble models; the first was based on the three models having the highest accuracy, and the second was based on the five models having the highest accuracy. The area under the curve (AUC) of the case that used the three most accurate models was the highest at 0.998. This study demonstrates that an ensemble of transfer-learning-based models can accurately classify and detect fractures, even in an imbalanced dataset. This artificial intelligence (AI)-assisted diagnostic application could support decision-making and reduce the workload of clinicians with its high speed and accuracy.

Clinical Validation of an Artificial Intelligence Model for Detecting Distal Radius, Ulnar Styloid, and Scaphoid Fractures on Conventional Wrist Radiographs

This study aimed to assess the feasibility and performance of an artificial intelligence (AI) model for detecting three common wrist fractures: distal radius, ulnar styloid process, and scaphoid. The AI model was trained with a dataset of 4432 images containing both fractured and non-fractured wrist images. In total, 593 subjects were included in the clinical test. Two human experts independently diagnosed and labeled the fracture sites using bounding boxes to build the ground truth. Two novice radiologists also performed the same task, both with and without model assistance. The sensitivity, specificity, accuracy, and area under the curve (AUC) were calculated for each wrist location. The AUC for detecting distal radius, ulnar styloid, and scaphoid fractures per wrist were 0.903 (95% C.I. 0.887-0.918), 0.925 (95% C.I. 0.911-0.939), and 0.808 (95% C.I. 0.748-0.967), respectively. When assisted by the AI model, the scaphoid fracture AUC of the two novice radiologists significantly increased from 0.75 (95% C.I. 0.66-0.83) to 0.85 (95% C.I. 0.77-0.93) and from 0.71 (95% C.I. 0.62-0.80) to 0.80 (95% C.I. 0.71-0.88), respectively. Overall, the developed AI model was found to be reliable for detecting wrist fractures, particularly for scaphoid fractures, which are commonly missed.

Detecting pediatric wrist fractures using deep-learning-based object detection

BACKGROUND

Missed fractures are the leading cause of diagnostic error in the emergency department, and fractures of pediatric bones, particularly subtle wrist fractures, can be misidentified because of their varying characteristics and responses to injury.

OBJECTIVE

This study evaluated the utility of an object detection deep learning framework for classifying pediatric wrist fractures as positive or negative for fracture, including subtle buckle fractures of the distal radius, and evaluated the performance of this algorithm as augmentation to trainee radiograph interpretation.

MATERIALS AND METHODS

We obtained 395 posteroanterior wrist radiographs from unique pediatric patients (65% positive for fracture, 30% positive for distal radial buckle fracture) and divided them into train (n = 229), tune (n = 41) and test (n = 125) sets. We trained a Faster R-CNN (region-based convolutional neural network) deep learning object-detection model. Two pediatric and two radiology residents evaluated radiographs initially without the artificial intelligence (AI) assistance, and then subsequently with access to the bounding box generated by the Faster R-CNN model.

RESULTS

The Faster R-CNN model demonstrated an area under the curve (AUC) of 0.92 (95% confidence interval [CI] 0.87-0.97), accuracy of 88% (n = 110/125; 95% CI 81-93%), sensitivity of 88% (n = 70/80; 95% CI 78-94%) and specificity of 89% (n = 40/45, 95% CI 76-96%) in identifying any fracture and identified 90% of buckle fractures (n = 35/39, 95% CI 76-97%). Access to Faster R-CNN model predictions significantly improved average resident accuracy from 80 to 93% in detecting any fracture (P < 0.001) and from 69 to 92% in detecting buckle fracture (P < 0.001). After accessing AI predictions, residents significantly outperformed AI in cases of disagreement (73% resident correct vs. 27% AI, P = 0.002).

CONCLUSION

An object-detection-based deep learning approach trained with only a few hundred examples identified radiographs containing pediatric wrist fractures with high accuracy. Access to model predictions significantly improved resident accuracy in diagnosing these fractures.

Artificial Intelligence in Fracture Detection: A Systematic Review and Meta-Analysis

Background Patients with fractures are a common emergency presentation and may be misdiagnosed at radiologic imaging. An increasing number of studies apply artificial intelligence (AI) techniques to fracture detection as an adjunct to clinician diagnosis. Purpose To perform a systematic review and meta-analysis comparing the diagnostic performance in fracture detection between AI and clinicians in peer-reviewed publications and the gray literature (ie, articles published on preprint repositories). Materials and Methods A search of multiple electronic databases between January 2018 and July 2020 (updated June 2021) was performed that included any primary research studies that developed and/or validated AI for the purposes of fracture detection at any imaging modality and excluded studies that evaluated image segmentation algorithms. Meta-analysis with a hierarchical model to calculate pooled sensitivity and specificity was used. Risk of bias was assessed by using a modified Prediction Model Study Risk of Bias Assessment Tool, or PROBAST, checklist. Results Included for analysis were 42 studies, with 115 contingency tables extracted from 32 studies (55 061 images). Thirty-seven studies identified fractures on radiographs and five studies identified fractures on CT images. For internal validation test sets, the pooled sensitivity was 92% (95% CI: 88, 93) for AI and 91% (95% CI: 85, 95) for clinicians, and the pooled specificity was 91% (95% CI: 88, 93) for AI and 92% (95% CI: 89, 92) for clinicians. For external validation test sets, the pooled sensitivity was 91% (95% CI: 84, 95) for AI and 94% (95% CI: 90, 96) for clinicians, and the pooled specificity was 91% (95% CI: 81, 95) for AI and 94% (95% CI: 91, 95) for clinicians. There were no statistically significant differences between clinician and AI performance. There were 22 of 42 (52%) studies that were judged to have high risk of bias. Meta-regression identified multiple sources of heterogeneity in the data, including risk of bias and fracture type. Conclusion Artificial intelligence (AI) and clinicians had comparable reported diagnostic performance in fracture detection, suggesting that AI technology holds promise as a diagnostic adjunct in future clinical practice. Clinical trial registration no. CRD42020186641 © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Cohen and McInnes in this issue.

A Progressive and Cross-Domain Deep Transfer Learning Framework for Wrist Fracture Detection

There has been an amplified focus on and benefit from the adoption of artificial intelligence (AI) in medical imaging applications. However, deep learning approaches involve training with massive amounts of annotated data in order to guarantee generalization and achieve high accuracies. Gathering and annotating large sets of training images require expertise which is both expensive and time-consuming, especially in the medical field. Furthermore, in health care systems where mistakes can have catastrophic consequences, there is a general mistrust in the black-box aspect of AI models. In this work, we focus on improving the performance of medical imaging applications when limited data is available while focusing on the interpretability aspect of the proposed AI model. This is achieved by employing a novel transfer learning framework, progressive transfer learning, an automated annotation technique and a correlation analysis experiment on the learned representations.

Progressive transfer learning helps jump-start the training of deep neural networks while improving the performance by gradually transferring knowledge from two source tasks into the target task. It is empirically tested on the wrist fracture detection application by first training a general radiology network RadiNet and using its weights to initialize RadiNetwrist , that is trained on wrist images to detect fractures. Experiments show that RadiNetwrist achieves an accuracy of 87% and an AUC ROC of 94% as opposed to 83% and 92% when it is pre-trained on the ImageNet dataset.

This improvement in performance is investigated within an explainable AI framework. More concretely, the learned deep representations of RadiNetwrist are compared to those learned by the baseline model by conducting a correlation analysis experiment. The results show that, when transfer learning is gradually applied, some features are learned earlier in the network. Moreover, the deep layers in the progressive transfer learning framework are shown to encode features that are not encountered when traditional transfer learning techniques are applied.

In addition to the empirical results, a clinical study is conducted and the performance of RadiNetwrist is compared to that of an expert radiologist. We found that RadiNetwrist exhibited similar performance to that of radiologists with more than 20 years of experience.

This motivates follow-up research to train on more data to feasibly surpass radiologists’ performance, and investigate the interpretability of AI models in the healthcare domain where the decision-making process needs to be credible and transparent.

Improving Radiographic Fracture Recognition Performance and Efficiency Using Artificial Intelligence

Background Missed fractures are a common cause of diagnostic discrepancy between initial radiographic interpretation and the final read by board-certified radiologists. Purpose To assess the effect of assistance by artificial intelligence (AI) on diagnostic performances of physicians for fractures on radiographs. Materials and Methods This retrospective diagnostic study used the multi-reader, multi-case methodology based on an external multicenter data set of 480 examinations with at least 60 examinations per body region (foot and ankle, knee and leg, hip and pelvis, hand and wrist, elbow and arm, shoulder and clavicle, rib cage, and thoracolumbar spine) between July 2020 and January 2021. Fracture prevalence was set at 50%. The ground truth was determined by two musculoskeletal radiologists, with discrepancies solved by a third. Twenty-four readers (radiologists, orthopedists, emergency physicians, physician assistants, rheumatologists, family physicians) were presented the whole validation data set (n = 480), with and without AI assistance, with a 1-month minimum washout period. The primary analysis had to demonstrate superiority of sensitivity per patient and the noninferiority of specificity per patient at -3% margin with AI aid. Stand-alone AI performance was also assessed using receiver operating characteristic curves. Results A total of 480 patients were included (mean age, 59 years ± 16 [standard deviation]; 327 women). The sensitivity per patient was 10.4% higher (95% CI: 6.9, 13.9; P < .001 for superiority) with AI aid (4331 of 5760 readings, 75.2%) than without AI (3732 of 5760 readings, 64.8%). The specificity per patient with AI aid (5504 of 5760 readings, 95.6%) was noninferior to that without AI aid (5217 of 5760 readings, 90.6%), with a difference of +5.0% (95% CI: +2.0, +8.0; P = .001 for noninferiority). AI shortened the average reading time by 6.3 seconds per examination (95% CI: -12.5, -0.1; P = .046). The sensitivity by patient gain was significant in all regions (+8.0% to +16.2%; P < .05) but shoulder and clavicle and spine (+4.2% and +2.6%; P = .12 and .52). Conclusion AI assistance improved the sensitivity and may even improve the specificity of fracture detection by radiologists and nonradiologists, without lengthening reading time. © RSNA, 2021 Online supplemental material is available for this article. See also the editorial by Link and Pedoia in this issue.

An increasing number of convolutional neural networks for fracture recognition and classification in orthopaedics : are these externally validated and ready for clinical application?

Aims

The number of convolutional neural networks (CNN) available for fracture detection and classification is rapidly increasing. External validation of a CNN on a temporally separate (separated by time) or geographically separate (separated by location) dataset is crucial to assess generalizability of the CNN before application to clinical practice in other institutions. We aimed to answer the following questions: are current CNNs for fracture recognition externally valid?; which methods are applied for external validation (EV)?; and, what are reported performances of the EV sets compared to the internal validation (IV) sets of these CNNs?

Methods

The PubMed and Embase databases were systematically searched from January 2010 to October 2020 according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. The type of EV, characteristics of the external dataset, and diagnostic performance characteristics on the IV and EV datasets were collected and compared. Quality assessment was conducted using a seven-item checklist based on a modified Methodologic Index for NOn-Randomized Studies instrument (MINORS).

Results

Out of 1,349 studies, 36 reported development of a CNN for fracture detection and/or classification. Of these, only four (11%) reported a form of EV. One study used temporal EV, one conducted both temporal and geographical EV, and two used geographical EV. When comparing the CNN’s performance on the IV set versus the EV set, the following were found: AUCs of 0.967 (IV) versus 0.975 (EV), 0.976 (IV) versus 0.985 to 0.992 (EV), 0.93 to 0.96 (IV) versus 0.80 to 0.89 (EV), and F1-scores of 0.856 to 0.863 (IV) versus 0.757 to 0.840 (EV).

Conclusion

The number of externally validated CNNs in orthopaedic trauma for fracture recognition is still scarce. This greatly limits the potential for transfer of these CNNs from the developing institute to another hospital to achieve similar diagnostic performance. We recommend the use of geographical EV and statements such as the Consolidated Standards of Reporting Trials–Artificial Intelligence (CONSORT-AI), the Standard Protocol Items: Recommendations for Interventional Trials–Artificial Intelligence (SPIRIT-AI) and the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis–Machine Learning (TRIPOD-ML) to critically appraise performance of CNNs and improve methodological rigor, quality of future models, and facilitate eventual implementation in clinical practice.

Cite this article: Bone Jt Open 2021;2(10):879–885.

How Can a Deep Learning Algorithm Improve Fracture Detection on X-rays in the Emergency Room?

The growing need for emergency imaging has greatly increased the number of conventional X-rays, particularly for traumatic injury. Deep learning (DL) algorithms could improve fracture screening by radiologists and emergency room (ER) physicians. We used an algorithm developed for the detection of appendicular skeleton fractures and evaluated its performance for detecting traumatic fractures on conventional X-rays in the ER, without the need for training on local data. This algorithm was tested on all patients (N = 125) consulting at the Louis Mourier ER in May 2019 for limb trauma. Patients were selected by two emergency physicians from the clinical database used in the ER. Their X-rays were exported and analyzed by a radiologist. The prediction made by the algorithm and the annotation made by the radiologist were compared. For the 125 patients included, 25 patients with a fracture were identified by the clinicians, 24 of whom were identified by the algorithm (sensitivity of 96%). The algorithm incorrectly predicted a fracture in 14 of the 100 patients without fractures (specificity of 86%). The negative predictive value was 98.85%. This study shows that DL algorithms are potentially valuable diagnostic tools for detecting fractures in the ER and could be used in the training of junior radiologists.