Ensemble Deep Convolutional Neural Network to Identify Fractured Limbs using CT Scans

Abstract

Accurate classification between fractured and intact bones in Computed Tomography (CT) scan serves as a precursor to further treatment planning. CNN is no exception to handle this, and as an example AlexNet ranked top in the ImageNet challenge (2012). To overcome generalization errors, we propose to ensemble deep convolutional neural networks to check how well fractured limbs can be analyzed. It primarily includes voting (soft and hard), stacking, bagging, and feature soup on a backbone consisting of VGG19, ResNet152, Inception, MobileNet, and DenseNet169. On a clinically annotated dataset of size 5,567 CT scans, we achieved the highest accuracy of 0.977, precision of 0.959, recall of 0.960, F1-score of 0.960, and AUC of 0.971. To the best of our knowledge, this is the first time this dataset has been used to classify fractured and intact bones.