Frequency-Enhanced Generative Adversarial Network for Unpaired Medical Image Translation

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Frequency-Enhanced Generative Adversarial Network for Unpaired
        Medical Image Translation"

By

BAE Juyoung

Abstract:

Unpaired image-to-image translation (UIIT) has a wide application in 
medical imaging for generating synthesized cross-domain images without the 
need for strictly paired dataset. Most current state-of-the-art (SOTA) 
UIIT approaches rely on spatial image representation, which has limited 
capability to enrich its own feature space without any complicated 
transformation algorithms. On the other hand, images represented in their 
frequency domain can readily be decomposed or converted into more 
sophisticated components that altogether yield a much richer set of 
features that represent the same image. To this end, we propose a novel 
generative adversarial network (GAN) -- based UIIT network that extensively 
utilizes the image frequency domain. Specifically, we introduce a 
cycle-consistent GAN with a discriminator that operates on a stack of 
different frequency components (FCs). Also, given that some FCs carry more 
discriminative information depending on the application, we incorporate a 
specially designed FC-wise attention module in our discriminator for 
extended flexibility to different medical applications. To illustrate the 
versatility of our approach, we conducted experiments on two unrelated 
UIIT applications with different cross-domain distributions. Experiments 
show that on both datasets, our model achieves superiorUIIT performance to 
the SOTA UIIT model.


Date            : 3 May 2022 (Tuesday)

Time            : 15:35-16:15

Zoom Link:
https://hkust.zoom.us/j/92797640941?pwd=ZWZGcU1aZzR4b1Z4V3dxLytoVTVSUT09

Meeting ID      : 927 9764 0941

Passcode        : csefyp

Advisor         : Dr. CHEN Hao

2nd Reader      : Dr. XU Dan