Effect of GAN augmented dataset size on deep learning-based ultrasound bone segmentation model training

Asaduz Zaman, Hyunhee Bang, Chul-woo Park, Ilhyung Park, Sanghyun Joung

Keywords: Ultrasound bone segmentation, GAN, data augmentation

TL;DR: Although the use of the GAN-augmented training dataset in addition to standard augmentation approaches helps perform the network better, the addition of multi-fold GAN-augmented datasets has no noticeable performance gain

Abstract: Recently, ultrasound imaging is increasingly being used as intra-operative imaging modality in osteolytic bone surgery due to its cost-effectiveness and radiation-free nature. Deep learning approaches have shown remarkable success in segmenting bone surface from ultrasound images. However, limited training dataset size has always hindered the success of deep learning approaches, which is especially evident in ultrasound bone segmentation since there is no publicly available quality dataset. To resolve the issue, in addition to standard data augmentation approaches, generative adversarial networks (GANs) have recently been used for generating augmented training samples. Although the addition of the generative approach in data augmentation is observed to have a positive effect on deep learning architecture's performance, the question about the effect of using a multi-fold of GAN generated training dataset remains unanswered. In this work, we have generated 14 fold of GAN-augmented training dataset and evaluated the performance of the network for successively increased dataset size. Our test results show that although the use of the GAN-augmented training dataset in addition to standard augmentation approaches helps perform the network better, the addition of multi-fold GAN-augmented datasets has no noticeable performance gain.

Track: short paper