Model architectures.

<div> Introduction The interpretation of plain hip radiographs can vary widely among physicians. This study aimed to develop and validate a deep learning-based screening model for distinguishing normal hips from severe hip diseases on plain radiographs. Methods Electronic medical records and plain radiograph from 2004 to 2012 were used to construct two patient groups: the hip disease group (those who underwent total hip arthroplasty) and normal group. A total of 1,726 radiographs (500 normal hip radiographs and 1,226 radiographs with hip diseases, respectively) were included and were allocated for training (320 and 783), validation (80 and 196), and test (100 and 247) groups. Four different models were designed–raw image for both training and test set, preprocessed image for training but raw image for the test set, preprocessed images for both sets, and change of backbone algorithm from DenseNet to EfficientNet. The deep learning models were compared in terms of accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, and area under the receiver operating characteristic curve (AUROC). Results The mean age of the patients was 54.0 ± 14.8 years in the hip disease group and 49.8 ± 14.9 years in the normal group. The final model showed highest performance in both the internal test set (accuracy 0.96, sensitivity 0.96, specificity 0.97, PPV 0.99, NPV 0.99, F1-score 0.97, and AUROC 0.99) and the external validation set (accuracy 0.94, sensitivity 0.93, specificity 0.96, PPV 0.95, NPV 0.93, F1-score 0.94, and AUROC 0.98). In the gradcam image, while the first model depended on unrelated marks of radiograph, the second and third model mainly focused on the femur shaft and sciatic notch, respectively. Conclusion The deep learning-based model showed high accuracy and reliability in screening hip diseases on plain radiographs, potentially aiding physicians in more accurately diagnosing hip conditions. </div>