One-Stage vs Two-Stage Deep Learning Method for Bone Abnormality Detection

This paper introduces an automatic reliable bone abnormality detection using two different classification approaches. The introduced method considers the seven extremity upper bones namely shoulder, humerus, forearm, elbow, wrist, hand, and finger. For this purpose, two different classification approaches are considered by utilizing Inception pre-trained model. The images are first enhanced by utilizing adaptive histogram equalization. Thereafter, the enhanced images are fed into the classification step. The first classification approach is a one-stage approach where the model takes x-ray images as input and outputs both the bone type and whether it is normal or not. While the second approach is a two-stage hierarchical approach where the x-ray images are fed into the first stage to classify the bone type and then fed into the second stage to detect the bone abnormality. The average sensitivity achieved is 61.53% and 70.04% and the average specificity achieved is 84.22% and 74.76% for one-stage and two-stage hierarchical approaches, respectively. All the experiments were carried out using MURA dataset.