Scholarlink

<p>Computer-assisted diagnostic tools may help improve the consistency and reliability of ultrasonographic screening for developmental dysplasia of the hip (DDH). This study aimed to compare preprocessing strategies for convolutional neural network (CNN)-based classification of infant hip ultrasound images under heterogeneous image quality conditions. We evaluated six approaches for classifying infant hip ultrasound images as normal or abnormal: original-image CNN classification, CycleGAN-based domain translation in two directions, DeepLabV3-Plus–based segmentation preprocessing, and two combined approaches using CycleGAN followed by DeepLabV3-Plus. Each approach was independently trained and evaluated ten times using ninefold cross-validation, and performance was assessed using the area under the receiver operating characteristic curve (AUC). The baseline approach using original images achieved a mean AUC of 0.724 (95% confidence interval [CI]: 0.711–0.737). CycleGAN-based domain translation did not improve performance, with mean AUCs of 0.719 (95% CI: 0.703–0.736) for Dataset A-to-B translation and 0.679 (95% CI: 0.658–0.700) for Dataset B-to-A translation. DeepLabV3-Plus–based segmentation preprocessing showed the numerically highest mean AUC of 0.791 (95% CI: 0.773–0.809), while the combined approaches achieved mean AUCs of 0.780 and 0.783. Pairwise comparisons with Bonferroni correction showed no statistically significant differences between approaches (all adjusted p-values &gt; 0.05). These exploratory findings suggest that segmentation-based preprocessing may be a promising strategy for CNN-based classification of infant hip ultrasound images under heterogeneous imaging conditions, whereas CycleGAN-based domain translation did not provide a measurable benefit in this setting. Further validation using larger multicenter datasets is required before clinical applicability can be established.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning for developmental dysplasia of the hip: comparative classification strategies for ultrasound imaging with heterogeneous image quality

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Shota Ichikawa,
  • Yohan Kondo

摘要

Computer-assisted diagnostic tools may help improve the consistency and reliability of ultrasonographic screening for developmental dysplasia of the hip (DDH). This study aimed to compare preprocessing strategies for convolutional neural network (CNN)-based classification of infant hip ultrasound images under heterogeneous image quality conditions. We evaluated six approaches for classifying infant hip ultrasound images as normal or abnormal: original-image CNN classification, CycleGAN-based domain translation in two directions, DeepLabV3-Plus–based segmentation preprocessing, and two combined approaches using CycleGAN followed by DeepLabV3-Plus. Each approach was independently trained and evaluated ten times using ninefold cross-validation, and performance was assessed using the area under the receiver operating characteristic curve (AUC). The baseline approach using original images achieved a mean AUC of 0.724 (95% confidence interval [CI]: 0.711–0.737). CycleGAN-based domain translation did not improve performance, with mean AUCs of 0.719 (95% CI: 0.703–0.736) for Dataset A-to-B translation and 0.679 (95% CI: 0.658–0.700) for Dataset B-to-A translation. DeepLabV3-Plus–based segmentation preprocessing showed the numerically highest mean AUC of 0.791 (95% CI: 0.773–0.809), while the combined approaches achieved mean AUCs of 0.780 and 0.783. Pairwise comparisons with Bonferroni correction showed no statistically significant differences between approaches (all adjusted p-values > 0.05). These exploratory findings suggest that segmentation-based preprocessing may be a promising strategy for CNN-based classification of infant hip ultrasound images under heterogeneous imaging conditions, whereas CycleGAN-based domain translation did not provide a measurable benefit in this setting. Further validation using larger multicenter datasets is required before clinical applicability can be established.