Automatic Keypoint Detection and Context Analysis of X-Ray

This study focuses on an automated, trainable processing pipeline enabling automatic, accurate detection of selected keypoints in hip X-ray. By using a hybrid loss function combining Cross-Entropy and Dice Loss, we address the challenges of class imbalance and improve the detection of structures with very small population representation. Auxiliary output of introduced method is regression head, which fuses embedding output with the output of the decoder in order to predict values of Leg Length Discrepancy and Leg Offset Discrenacy of a patient.