Improving GAN learning dynamics for thyroid nodule segmentation | |
| Author | Alisa Kunapinun |
| Call Number | AIT Diss. no.ISE-22-01 |
| Subject(s) | Thyroid Nodule Deep learning |
| Note | A dissertation submitted in partial fulfillment of the requirements for the degree of Doctoral of Engineering of Engineering in Mechatronics |
| Publisher | Asian Institute of Technology |
| Abstract | The thyroid gland, which is responsible for secretion of critical hormones that regulate the body, is susceptible to a number of pathological conditions best diagnosed using ultrasound. A common diagnosis method involves identifying then characterizing the appearance of thyroid nodules: solid or fluid-filled lumps that could be benign or malignant. Physicians could be aided in this endeavor by an automated system to precisely identify nodules in a given ultrasound image. This thesis therefore presents a novel algorithm for segmenting thyroid nodules in ultrasound images named StableSeg GAN. The algorithm is based on the concept of image-to-image translation, which combines traditional supervised semantic segmentation with unsupervised learning using generative adversarial networks (GANs). GANs have been found to improve semantic segmentation models’ performance in specific tasks. However, GAN learning dynamics are famously unstable, oftentimes leading to mode collapse. It is well known that controlling the discriminator in a GAN to not learn too quickly often improves generator learning, making the learning smoother and avoiding mode collapse. StableSeg GANs exploit the concept of closed-loop control of the gain on the loss output of the discriminator to stabilize training. We find that gain control leads to smoother generator training and avoids the mode collapse that typically occurs when the discriminator learns too quickly relative to the generator. We also find that the combination of the supervised and unsupervised learning styles encourages both low level accuracy and high-level consistency. As a test of the concept of controlled hybrid supervised and unsupervised semantic segmentation, StableSeg GANs use DeeplabV3+ as the generator, Resnet18 as the discriminator, and PID control to stabilize the GAN learning process. The new model is superior to the state-of-the-art DeeplabV3+ in terms of intersection over union (IoU), with a mean IoU of 81.26% over a challenging test set. The results of our thyroid nodule segmentation experiments show that StableSeg GANs have flexibility to segment nodules more accurately than supervised segmentation models or uncontrolled GANs. |
| Year | 2022 |
| Type | Dissertation |
| School | School of Engineering and Technology (SET) |
| Department | Department of Industrial Systems Engineering (DISE) |
| Academic Program/FoS | Industrial Systems Engineering (ISE) |
| Chairperson(s) | Mongkol Ekpanyapong;Dittapong Songsaeng (Co-Chairperson); |
| Examination Committee(s) | Siridech Boonsang;Dailey, Matthew N.;Manukid Parnichkun;Chadaporn Keatmanee; |
| Scholarship Donor(s) | Royal Thai Government; |
| Degree | Thesis (Ph. D.) - Asian Institute of Technology, 2022 |