University of Twente Student Theses

Login

The usability of generative adversarial networks for automatic segmentation of lung nodules in CT-images

Elst, S. van (2021) The usability of generative adversarial networks for automatic segmentation of lung nodules in CT-images.

[img] PDF
5MB
Abstract:Lung cancer is a highly prevailing disease and early detection and treatment is crucial to increase the likelihood of survival. In many lung cancer procedures, the segmentation of lung nodules in CT-images is an essential step. However, manual annotation of the nodules is a difficult and timeconsuming task, which relies heavily on the experience of a radiologist. To assist radiologists in this process, computer-assisted segmentation systems could be a promising tool. The aim of this study was to explore the feasibility of applying generative adversarial networks (GANs) to automatically segment lung nodules from entire 2D CT-images. The network architecture proposed in this thesis was designed to address three commonly occurring segmentation challenges: (i) variability in lung nodule appearance, (ii) class imbalance, and (iii) GAN training instability. The overall network followed the structure of a conditional image-to-image translation GAN, in which a U-Net was used as the generator network. To alleviate the aforementioned challenges, the following measures were adopted: atrous spatial pyramid pooling (ASPP) modules to reduce the influence of appearance variability, an additional Dice loss to counteract the class imbalance and a multiscale-L1 critic for adversarial training stability. The added value of each module to the generator network was evaluated in an ablation study. Additionally, the usefulness of GANs compared to the state-of-the-art segmentation network, the U-Net, was explored. The results of the ablation study showed that the addition of ASPP modules to various locations of the generator network achieved inferior or comparable results to the results of the generator without ASPP modules. The Dice loss appeared to be an essential addition to produce accurate segmentation results for both the U-Net and the GAN. The GAN with a multiscale-L1 critic did show a stable training course, in contrast to more conventional GANs. However, introducing adversarial training by adding a critic did not improve the performance of the generator network alone. Overall, the GAN-based method obtained worse results than the U-Net, whereas the U-Net itself achieved excellent, state-of-the-art segmentation results compared to other works. These results suggest that the Dice loss and the multiscale-L1 critic did facilitate in solving the segmentation challenges, whereas the ASPP modules were not of added value. Although the GAN with an additional Dice loss showed sufficient performance and comparable results to other GAN-based methods, it could not outperform the U-Net. In conclusion, this study showed that GANs can be applied for lung nodule segmentation in entire 2D CT-images, but the proposed GAN-architecture is not advantageous over the current state-of-the-art segmentation methods for this application. Future work could focus on expanding and balancing the dataset, using more suitable loss functions or extending the 2D models into 3D models.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:31 mathematics, 44 medicine, 50 technical science in general
Programme:Biomedical Engineering MSc (66226)
Link to this item:https://purl.utwente.nl/essays/86203
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page