Automated Abnormality Diagnosis in Musculoskeletal Radiographs Using a Two‐Stage <scp>CNN</scp> Method

ABSTRACT Musculoskeletal disorders (MDs) consume a larger component of health expenditure and these are on the verge of becoming the leading contributor to worldwide disability. Accurately classifying musculoskeletal radiographs is often time‐consuming, error‐prone, and requires experienced radiologists. Radiologists across subspecialties could greatly benefit from deep learning based automated tools in accurately diagnosing musculoskeletal abnormalities. This paper proposes a novel two‐stage body part identification and abnormality diagnosis system for the upper extremity bone X‐ray images. The first stage of the proposed method uses the customized DenseNet201 classifier for bone type identification, whereas the second stage uses two customized DenseNet201 classifiers to help diagnose corresponding bone abnormalities. The first stage classifier achieved the highest precision and recall scores of 1.0 for both the finger and shoulder images, whereas the classifier achieved the highest F1‐score of 0.99 for the shoulder images. The first DenseNet201 classifier of the second stage achieved the highest AUC‐ROC and F1‐scores of 0.87 and 0.87, respectively, for both humerus and elbow images, whereas the classifier achieved the highest Cohen's kappa score of 0.75 for the elbow images. The other classifier of the second stage achieved the highest AUC‐ROC and Cohen's kappa scores of 0.82 and 0.64, respectively, for the forearm images, whereas the highest F1‐score of 0.79 was achieved for both shoulder and forearm images. The proposed method benefited from the divide‐and‐conquer strategy, improving F1‐scores by 2.33%, 3.90%, 3.7%, and 7.95%, for the normal images of the elbow, finger, hand, and wrist. Likewise, for abnormal images of the elbow, finger, humerus, and shoulder, the method improved F1‐scores by 4.82%, 8.11%, 1.16%, and 1.28%, while utilizing only two classifiers in the second stage, compared to seven in the existing benchmark method. These results demonstrate that the proposed system achieves competitive performance and holds strong potential for real‐world clinical applications in musculoskeletal imaging.