City Research Online

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.

Publication Type:	Conference or Workshop Item (Paper)
Additional Information:	© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Publisher Keywords:	CNN, X-ray, Wrist, Classification
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > RC Internal medicine R Medicine > RC Internal medicine > RC0321 Neuroscience. Biological psychiatry. Neuropsychiatry
Departments:	School of Science & Technology > Computer Science School of Science & Technology > Computer Science > giCentre

Export

Downloads