Impact of data augmentation techniques on a deep learning based medical imaging task

摘要

For any deep learning (DL) based task, model generalization and prediction performance improve as a function of trainingdata set size and variety. However, its application to medical imaging is still challenging because of the limited availabilityof high-quality and sufficiently diverse annotated data. Data augmentation techniques can improve the model performancewhen the available dataset size is limited. Anatomy region localization from medical images can be automated with deeplearning and is important for tasks such as organ segmentation and lesion detection. Different data augmentation methodswere compared for DL based anatomy region localization with computed tomography images. The impact of differentneural network architectures was also explored. The prediction accuracy on an independent test set improved from 88%to 97% with optimal selection of data augmentation and architecture while using the same training dataset. Dataaugmentation steps such as zoom, translation and flips had incremental effect on classifier performance whereas samplewisemean shift appeared to degrade the classifier performance. Global average pooling improved classifier accuracycompared to fully-connected layer when limited data augmentation was used. All model architectures converged to anoptimal performance with the right combination of augmentation steps. Prediction inaccuracies were mostly observed inthe boundary regions between anatomies. The networks also successfully localized anatomy for Positron EmissionTomography studies reaching an accuracy of up to 97%. Similar impact of data augmentation and pooling layer was alsoobserved.