Radiography Classification: A Comparison between Eleven Convolutional Neural Networks

Abstract

This paper investigates the classification of normal and abnormal radiographic images. Eleven convolutional neural network architectures (GoogleNet, Vgg-19, AlexNet, SqueezeNet, ResNet-18, Inception-v3, ResNet-50, Vgg-16, ResNet-101, DenseNet-201 and Inception-ResNet-v2) were used to classify a series of x-ray images from Stanford Musculoskeletal Radiographs (MURA) dataset corresponding to the wrist images of the data base. For each architecture, the results were compared against the known labels (normal / abnormal) and then the following metrics were calculated: accuracy (labels correctly classified) and Cohen's kappa (a measure of agreement) following MURA guidelines. Numerous experiments were conducted by changing classifiers (Adam, Sgdm, RmsProp), the number of epochs, with/without data augmentation. The best results were provided by InceptionResnet-v2 (Mean accuracy = 0.723, Mean Kappa = 0.506). Interestingly, these results lower than those reported in the Leaderboard of MURA. We speculate that to improve the results from basic CNN architectures several options could be tested, for instance: pre-processing, post-processing or domain knowledge, and ensembles.