DENOISING AND SUPER-RESOLUTION OF MEDICAL IMAGES BY WEAKLY AND SELF-SUPERVISED LEARNING

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


PhD Thesis Defence


Title: "DENOISING AND SUPER-RESOLUTION OF MEDICAL IMAGES BY WEAKLY AND
SELF-SUPERVISED LEARNING"

By

Mr. Siu Chung TSANG


Abstract:

Medical images usually suffer from high noise and limited resolution. In
many medical image modalities, the noise level and resolution are restricted
by various reasons, such as the integrity of the sample, subject movement,
signal interference, scanning time, hardware settings, and the list goes on.
Post-processing applications such as segmentation, diagnosis, and detection
rely on a high-quality input image. The doctor's decision on treatment
planning or biologist's research study also depends on the noise level and
resolution of an image.

Deep learning-based denoising and super-resolution models today have shown
encouraging results in both natural images and medical images. A major
limitation in the literature is that these two tasks are addressed
separately. Recently, studies have shown that the joint denoising and
super-resolution (JDSR) approach outperforms the sequential application of
the denoiser and super-resolution model. The training process of these
methods requires noise-free ground truth or multiple noisy captures.
However, these extra training data are often unavailable in many medical
image applications.

This manuscript proposes a new weakly-supervised method in which, different
from other approaches, the JDSR model is trained with a single noisy-HR
capture alone. We further introduce a novel blind-spots framework to
approximate the supervised approach. We present both theoretical explanation
and experimental analysis for our method validation.

Next, we go one step further to perform JDSR without any training data. On
top of that, many techniques described in existing literature require a
predetermined scale, often set at 2, 4, or 8. This constraint presents a
significant hurdle for biologists who need the flexibility to zoom in and
out when analyzing cells, particularly for fluorescence microscopy
applications. Therefore, by incorporating concepts from the diffusion model,
we address the JDSR task without relying on training data or a fixed scale.
Our proposed solution, the Continuous Diffusion Model, merges the diffusion
model with a well-designed encoder and decoder. This model performs the
diffusion process in continuous rather than discrete pixel space, lowering
computational costs and enabling high-quality image reconstruction at
arbitrary resolutions.


Date:			Monday, 26 June 2023

Time:			2:30am - 4:30pm

Venue:			Room 4475
			Lifts 25/26

Chairman:		Prof. Xueqing ZHANG (CIVL)

Committee Members:	Prof. Pedro SANDER (Supervisor)
			Prof. Albert CHUNG (Supervisor)
			Prof. Shing Chi CHEUNG
			Prof. Chiew Lan TAI
			Prof. Tsz Wai WONG (CBE)
			Prof. Ed X. WU (HKU)


**** ALL are Welcome ****