Objectives Approximately 6.9% of children in the United Kingdom have suffered physical abuse. Fractures are a common sign and must not be overlooked due to high recurrence and mortality rates. We aimed to train and assess the diagnostic accuracy of a deep learning-based artificial intelligence model (BoneView) in detecting inflicted fractures. Materials and methods This pragmatic retrospective diagnostic accuracy pilot study focuses on children under 5 years old who underwent skeletal survey examinations for suspected physical abuse at a single tertiary centre between 1st January 2000 and 31st December 2023. Radiographs were extracted from the Picture Archiving and Communication System and divided to retrain and test the model. Radiology reports and retrospective review by one observer were used as the reference standard. Results Our total dataset included 1740 patients (mean age, 8.77 months ± 8.343 [standard deviation], 1026 males). The model’s baseline performance recorded an area under the receiver operating curve (AUC) of 0.46 (95% CI: 0.38, 0.57), with a sensitivity of 44% (95% CI: 35%, 58%) and a specificity of 61% (95% CI: 52%, 71%). For preliminary model training, 329 of 1227 positive studies were annotated, yielding a revised AUC of 0.55 (95% CI: 0.48, 0.66), sensitivity of 52% (95% CI: 43%, 64%), and specificity of 67% (95% CI: 58%, 78%). Conclusion Preliminary training of a novel AI tool for detecting inflicted fractures yielded improved results from baseline performance. This justifies the completion of annotation and further training of this AI tool to potentially achieve clinically acceptable performance. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Double reporting of skeletal surveys is vital for identifying fractures caused by physical abuse, but some departments lack the expertise to double report these investigations. <Emphasis Type="BoldItalic">Findings</Emphasis> Preliminary retraining of a commercially available deep learning algorithm using radiographic skeletal surveys led to improved inflicted fracture detection accuracy. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> Training this deep learning algorithm using relevant imaging enhances its performance. An accurate tool for automated skeletal survey interpretation may improve outcomes for physically abused children by offering an additional diagnostic opinion. Graphical Abstract

Developing an artificial intelligence tool for detecting fractures of child abuse: preliminary findings

期刊论文
出版模式：

开放获取
发表日期：

2026年4月4日

查看全文

Samuel Evans,
Nihal Chanian,
Esther Bezzina,
Amaka C. Offiah

摘要

Objectives

Approximately 6.9% of children in the United Kingdom have suffered physical abuse. Fractures are a common sign and must not be overlooked due to high recurrence and mortality rates. We aimed to train and assess the diagnostic accuracy of a deep learning-based artificial intelligence model (BoneView) in detecting inflicted fractures.

Materials and methods

This pragmatic retrospective diagnostic accuracy pilot study focuses on children under 5 years old who underwent skeletal survey examinations for suspected physical abuse at a single tertiary centre between 1st January 2000 and 31st December 2023. Radiographs were extracted from the Picture Archiving and Communication System and divided to retrain and test the model. Radiology reports and retrospective review by one observer were used as the reference standard.

Results

Our total dataset included 1740 patients (mean age, 8.77 months ± 8.343 [standard deviation], 1026 males). The model’s baseline performance recorded an area under the receiver operating curve (AUC) of 0.46 (95% CI: 0.38, 0.57), with a sensitivity of 44% (95% CI: 35%, 58%) and a specificity of 61% (95% CI: 52%, 71%). For preliminary model training, 329 of 1227 positive studies were annotated, yielding a revised AUC of 0.55 (95% CI: 0.48, 0.66), sensitivity of 52% (95% CI: 43%, 64%), and specificity of 67% (95% CI: 58%, 78%).

Conclusion

Preliminary training of a novel AI tool for detecting inflicted fractures yielded improved results from baseline performance. This justifies the completion of annotation and further training of this AI tool to potentially achieve clinically acceptable performance.

Key Points

Question Double reporting of skeletal surveys is vital for identifying fractures caused by physical abuse, but some departments lack the expertise to double report these investigations.

Findings Preliminary retraining of a commercially available deep learning algorithm using radiographic skeletal surveys led to improved inflicted fracture detection accuracy.

Clinical relevance Training this deep learning algorithm using relevant imaging enhances its performance. An accurate tool for automated skeletal survey interpretation may improve outcomes for physically abused children by offering an additional diagnostic opinion.

Graphical Abstract