|
Traditional landmark labeling relies on manual labeling by physicians based on lower limb x-rays, which has problems such as high subjective error, low efficiency, and repetitive operation. To address the above issues, this paper proposes an improved HR-Net deep learning network for the automatic annotation of feature points. Automatic identification and labeling of hip center points, Fujisawa points, hinge points, and ankle center points in lower limb x-rays, and calculation of the Miniaci angle of correction. Using HR-Net as a basic framework to construct an automatic feature point detection model. The introduction of the dense connection mechanism improves the continuity of feature delivery; the ECA attention mechanism was used to assign weighting information at the channel level, enhancing the ability to extract feature information from common regions. Use the doctor's manual labeling results as the true value. Five objective evaluation metrics, including mean squared error, mean absolute error, interclass correlation coefficient, Bland-Altman plot, and Scatter plot, were reused to assess the performance of the improved HR-Net network and the consistency between the annotated results and the manual annotations. The results show that the improved HR-Net network has significantly improved accuracy in detecting feature points. The mean squared errors measured for the four feature points were 1.3428, 1.4940, 1.9125, and 1.4441; compared to the original HR-Net network, the mean squared error reduction was 47.72%, 43.45%, 30.33%, and 40.83% respectively. The mean absolute error of the measured Miniaci angle was 0.224°, with an intergroup correlation coefficient of 0.995. The model's measurement results showed good agreement with manually measured results by clinicians, indicating that this method can be used to assist clinicians in measuring and diagnosing correction angles for lower limb deformities.
|
