Deep learning technology has rapidly developed, and neural networks have been widely applied in the field of medical image processing. The measurement of hip-knee-ankle angle (HKAA) in X-ray images is of great significance for diagnosing joint diseases, evaluating surgical outcomes, and formulating treatment plans. HKAA is an important indicator for assessing the lower limb skeletal structure. In previous studies, some measurement methods involved two stages: object detection and keypoint detection. However, we propose a single-stage measurement method using keypoint detection. We applied a multi-classification U2-Net model to predict the keypoint regions of the hip, knee, and ankle in X-ray images, and used the centers of these three regions and the cosine law to determine the HKAA. In the experiment, we selected 200 full-length lower limb X-ray images provided by a hospital, annotated the keypoint locations with reference to orthopedic doctors, and created a dataset. Then, modifications were made to the U2-Net model, transforming it from a binary object detection model to a multi-classification object detection model. The model was trained and tested, and the consistency between the angles measured by the model and the annotated angles was evaluated. The experimental results show that the mean difference of the angles is 0.152° ± 0.244°, and the intraclass correlation coefficient is 0.989, indicating good consistency of the results. Compared to U-Net and other neural networks, this study achieves comparable results with a small number of samples and can provide more accurate predictions of the HKAA.