Deep Learning based Determination of Graf Standart Plane on Hip Ultrasound Scans

Graf's ultrasonography (US) method is one of the most commonly used imaging techniques for developmental dysplasia of the hip (DDH) and is universally accepted for the assessment of neonatal hips [1]. However, the training process is lengthy and requires supervision until the evaluator achieves expertise. Computer-based segmentation and object detection tools may assist less experienced evaluators in identifying anatomical structures and classifying hip US images. This method involves anatomical description as well as measuring bone and soft tissue coverage in coronal two-dimensional (2D) US images of the hip. During scanning with the ultrasound probe, the physician has to decide whether the image is in the standard plane and whether the image is measurable. An image must contain a straight iliac wing, lower limb of the ilium, and the labrum to be classified as measurable [2, 3]. Graf's method is prone to interpreter variability due to the anatomical complexity of the hip structures, which can lead to misclassification [4]. When anatomical regions are not precisely identified, the selection of points for angle calculations may not be accurately determined, rendering the image unacceptable for measurements. This study comparatively measured success using a different model of the YOLOv8 algorithm to detect the labrum, lower limb of the ilium, and iliac wing regions in 200 measurable hip ultrasonography images obtained in the standard plane. With the YOLOv8x configuration, the labrum, ilium, and acetabulum were detected with success rates of 93.57%, 98.30% and 94.25% respectively, with an intersection over union (IoU) of 0.25. Our findings indicate that the YOLOv8x-based algorithm shows significant promise for the detection of labrum, ilium, and iliac wing regions in the standard plane.