Automated Diagnosis of Lower Limb Alignment Deformities Using Deep Learning and Joint Orientation Analysis

The misalignment of the lower limbs may cause serious musculoskeletal issues when not discovered at the initial stage, but the traditional methods of diagnosis are subjective, time‐consuming, and rely on clinical experience. In solving these problems, this paper will present an automated convolutional neural network (CNN) to identify the presence of lower limb misalignments in X‐ray images. The anatomical landmarks have been identified to be complex, image quality has been inconsistent, and limited datasets with labels have historically obstructed consistent diagnosis. To overcome these difficulties, the proposed CNN recognizes the most important landmarks, i.e., the hip, knee, and ankle joints, and the data augmentation methods made it stronger and flexible to various imaging characteristics. The comparative analysis showed that the CNN had a high accuracy of 94.11, precision of 1.00, recall of 0.91666, F1‐score of 0.9565, and an AUC of 0.97, which is significantly better than the traditional models such as support vector machine (SVM), random forest (RF), logistic regression, and fully connected networks (FCNs). Statistical analysis confirmed these performance differences were significant, with paired t ‐tests showing p values well below 0.05 ( p = 0.004932 for SVM, p < 0.000001 for RF, p = 0.000197 for logistic regression, and p = 0.000031 for FCN). The confusion matrix analysis revealed exceptional performance with 44 true positives (TP), 0 false positives (FP), and only 4 false negatives (FN), demonstrating the model’s proficiency in accurately identifying misalignment cases while preventing misclassification of properly aligned images. Furthermore, 10‐fold cross‐validation confirmed the stability and generalization capacity of the CNN, yielding a mean accuracy of 93.98% ± 0.0124, flawless precision of 1.00 ± 0.00, strong recall of 0.9168 ± 0.0201, F1‐score of 0.9565 ± 0.0128, and consistently high AUC values of 0.9662 ± 0.0107 across folds. These outcomes emphasize the reliability of the model in minimizing both FP and FN, making it a dependable tool for clinical applications.