Objectives To evaluate how different test set sampling strategies—random selection and balanced sampling—affect the performance of artificial intelligence (AI) models in pediatric wrist fracture detection using radiographs, aiming to highlight the need for standardization in test set design. Materials and methods This retrospective study utilized the open-sourced GRAZPEDWRI-DX dataset of 6091 pediatric wrist radiographs. Two test sets, each containing 4588 images, were constructed: one using a balanced approach based on case difficulty, projection type, and fracture presence and the other a random selection. EfficientNet and YOLOv11 models were trained and validated on 18,762 radiographs and tested on both sets. Binary classification and object detection tasks were evaluated using metrics such as precision, recall, F1 score, AP50, and AP50-95. Statistical comparisons between test sets were performed using nonparametric tests. Results Performance metrics significantly decreased in the balanced test set with more challenging cases. For example, the precision for YOLOv11 models decreased from 0.95 in the random set to 0.83 in the balanced set. Similar trends were observed for recall, accuracy, and F1 score, indicating that models trained on easy-to-recognize cases performed poorly on more complex ones. These results were consistent across all model variants tested. Conclusion AI models for pediatric wrist fracture detection exhibit reduced performance when tested on balanced datasets containing more difficult cases, compared to randomly selected cases. This highlights the importance of constructing representative and standardized test sets that account for clinical complexity to ensure robust AI performance in real-world settings. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Do different sampling strategies based on samples’ complexity have an influence in deep learning models’ performance in fracture detection? <Emphasis Type="BoldItalic">Findings</Emphasis> AI performance in pediatric wrist fracture detection significantly drops when tested on balanced datasets with more challenging cases, compared to randomly selected cases. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> Without standardized and validated test datasets for AI that reflect clinical complexities, performance metrics may be overestimated, limiting the utility of AI in real-world settings. Graphical Abstract

Impact of test set composition on AI performance in pediatric wrist fracture detection in X-rays

期刊论文
出版模式：

开放获取
发表日期：

2025年5月16日

查看全文

Tristan Till,
Mario Scherkl,
Nikolaus Stranger,
Georg Singer,
Saskia Hankel,
Christina Flucher,
Franko Hržić,
Ivan Štajduhar,
Sebastian Tschauner

摘要

Objectives

To evaluate how different test set sampling strategies—random selection and balanced sampling—affect the performance of artificial intelligence (AI) models in pediatric wrist fracture detection using radiographs, aiming to highlight the need for standardization in test set design.

Materials and methods

This retrospective study utilized the open-sourced GRAZPEDWRI-DX dataset of 6091 pediatric wrist radiographs. Two test sets, each containing 4588 images, were constructed: one using a balanced approach based on case difficulty, projection type, and fracture presence and the other a random selection. EfficientNet and YOLOv11 models were trained and validated on 18,762 radiographs and tested on both sets. Binary classification and object detection tasks were evaluated using metrics such as precision, recall, F1 score, AP50, and AP50-95. Statistical comparisons between test sets were performed using nonparametric tests.

Results

Performance metrics significantly decreased in the balanced test set with more challenging cases. For example, the precision for YOLOv11 models decreased from 0.95 in the random set to 0.83 in the balanced set. Similar trends were observed for recall, accuracy, and F1 score, indicating that models trained on easy-to-recognize cases performed poorly on more complex ones. These results were consistent across all model variants tested.

Conclusion

AI models for pediatric wrist fracture detection exhibit reduced performance when tested on balanced datasets containing more difficult cases, compared to randomly selected cases. This highlights the importance of constructing representative and standardized test sets that account for clinical complexity to ensure robust AI performance in real-world settings.

Key Points

Question Do different sampling strategies based on samples’ complexity have an influence in deep learning models’ performance in fracture detection?

Findings AI performance in pediatric wrist fracture detection significantly drops when tested on balanced datasets with more challenging cases, compared to randomly selected cases.

Clinical relevance Without standardized and validated test datasets for AI that reflect clinical complexities, performance metrics may be overestimated, limiting the utility of AI in real-world settings.

Graphical Abstract