Improving GAN Learning Dynamics for Thyroid Nodule Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Thyroid nodules require accurate diagnosis and follow-up for effective patient management.
• Automated tools for nodule detection and segmentation can aid physicians in diagnosis and malignancy assessment.
• Existing segmentation models face challenges with accuracy and consistency, particularly in complex ultrasound images.

Purpose of the Study:

• To develop a novel algorithm for segmenting thyroid nodules in ultrasound images.
• To improve the performance of semantic segmentation models by combining supervised and unsupervised learning techniques.
• To address the instability and mode collapse issues common in Generative Adversarial Networks (GANs).

Main Methods:

• Developed a hybrid algorithm combining supervised semantic segmentation with unsupervised learning using GANs.
• Introduced a closed-loop control mechanism (PID control) to stabilize GAN training and prevent mode collapse.
• Utilized DeeplabV3+ as the generator and Resnet18 as the discriminator within the StableSeg GAN framework.

Main Results:

• The StableSeg GAN model demonstrated improved performance over traditional supervised models like DeeplabV3+.
• Achieved a mean Intersection over Union (IoU) of 81.26% on a challenging thyroid nodule dataset.
• The controlled hybrid approach resulted in smoother generator training and avoided discriminator-induced mode collapse.

Conclusions:

• StableSeg GAN offers a flexible and accurate solution for thyroid nodule segmentation in ultrasound images.
• The integration of controlled GANs enhances both low-level accuracy and high-level consistency in segmentation.
• This approach shows significant potential for advancing automated diagnostic tools in medical imaging.