Computer-aided diagnosis of DDH using ultrasound: deep learning for segmentation and accurate angle measurement aligned with radiologist's clinical workflow.
Authors
Affiliations (3)
Affiliations (3)
- Electrical and Electronics Engineering Department, Ankara University, Ankara.
- Electrical and Electronics Engineering Department, Ankara University, Turkish Accelerator and Radiation Laboratory (TARLA), Ankara. [email protected].
- Computer Engineering Department, Ankara University, Ankara.
Abstract
A computer-aided diagnosis (CAD) system for automated evaluation of developmental dysplasia of the hip (DDH) via ultrasound, integrating Deep Learning (DL) for anatomical segmentation and performing α&β angle calculations utilizing the Graf Method is presented. A custom image processing method excludes the inferior ilium's curvature during the baseline definition, enhancing accuracy and replicating radiologists' real-world workflow. Our dataset comprised 452 raw images from 370 newborns. For {'validation'+"test"}, {'nv=91'+"nte=45"}≡136 images were reserved (never augmented). Remaining 316 images were augmented to ntr=632 with (0%↔25%) random brightness manipulation for training. Totally (632+136)=768 images were annotated and split with the following true numbers and percentage: {'train',"validation",test}≡{'632',"91",45}≡{'82%',"12%",6%}. U-Net, MaskR-CNN, YOLOv8 and YOLOv11 were used for segmentation. α&β were measured using Method-I (centroid/orientation) and Method-II (Hough transform). An extended set of performance metrics-Precision, Recall, IoU, Dice, mAP-was calculated. Bland-Altman and Intraclass Correlation Coefficient (ICC) analyses compared CAD outputs with expert measurements. YOLOv11 showed the best segmentation performance (Precision:0.990, Recall:0.993, IoU:0.983, Dice:0.990, mAP:0.991). {ICCα, ICCβ} calculated using Method-I and Method-II were {0.895, 0.907} and {0.929, 0.952}, respectively, with Method-II outperforming Method-I. A clinically-aligned-CAD-system that integrates anatomical segmentation and α&β measurement-a combination rarely addressed in literature is introduced. By providing a comprehensive and standardized set of metrics, this work overcomes a common bottleneck in DL studies, namely heterogeneity in metric reporting, enabling better cross-study comparisons. Following curvature exclusion, obtained ICCs outperformed previous studies, demonstrating improved inter-rater reliability and strong agreement with expert radiologists, offering both technical robustness and clinical applicability in DDH assessment.