Candidate: Zhongling Wang
Title: Image Quality Assessment and Refocusing with Applications in Whole Slide Imaging
Date: May 31, 2024
Time: 1:00 PM
Place: REMOTE ATTENDANCE
Supervisor(s): Wang, Zhou
Abstract:
In the rapidly evolving field of general digital imaging and whole slide imaging, Image Quality Assessment (IQA) plays a crucial role in determining the perceptual quality of images and guiding image restoration. State-of-the-art IQA models are computationally expensive due to the use of complex deep learning architectures. The high computational cost poses a significant challenge in high-throughput Whole Slide Image (WSI) scanning platforms, which is both time-sensitive and power-limited. Moreover, most IQA models, while varied in design, often exhibit biases towards specific types of image content or distortions, a consequence of their underlying design principles or training data. To improve the quality of WSIs, we need to address the defocus problem, which is the most common distortion for a WSI. The transparency and uneven surface of tissue samples further complicate the restoration process for methods that lack an understanding of the 3D tissue structure. These issues emphasize the limitations and challenges faced by existing IQA and restoration models. This thesis proposes three novel and flexible approaches to mitigate these problems.
Addressing the efficiency concerns in whole slide imaging, this thesis presents a highly efficient model for Focus Quality Assessment (FQA), which is crucial for identifying out-of-focus artifacts in high-throughput WSI scanning. Different from photographic images, WSIs have much bigger dimensions, making most deep-learning based FQA models computationally infeasible. By improving efficiency, our model significantly reduces computational demands without compromising accuracy. This model can also be seamlessly integrated into and benefit lots of WSI restoration models. Furthermore, we introduce the first open-source, expert annotated FQA dataset, offering a comprehensive platform for developing and evaluating new FQA models.
To address the bias issue, we explore a novel approach of leveraging the strengths and mitigating the weaknesses of individual IQA models by fusing their scores, resulting in a more robust model. Unlike supervised learning-based fusion models, which can introduce new biases, our method employs an unsupervised approach for IQA score fusion using deep Maximum a Posteriori (MAP) estimation. Through fine-grained uncertainty estimation at the score level, the proposed model not only improves the accuracy of predictions but also demonstrates the unique ability to identify and reject inadequate models during the fusion process, as evidenced by comprehensive experimental validations. It not only introduces no additional biases, but it is also compatible with any combination of IQA datasets due to its unsupervised nature, which is impossible for supervised learning-based fusion models. We demonstrate its flexibility and robustness for both natural images and WSIs.
Addressing the challenge of defocus in WSI, traditional deblurring techniques often fail due to their inability to understand the 3D structure of tissue, which is not only transparent but also has an uneven surface. To overcome this limitation, we introduce the first virtual refocusing model tailored for WSI. Due to its implicit understanding of the 3D tissue structure, it can continuously adjust the focus level of each pixel of a single image captured at any focus plane. It is also designed to accept an arbitrary number of inputs, thereby integrating the complementary information of different focus levels. Moreover, we propose the first perceptual distance metric specifically developed for WSI, demonstrating its effectiveness as a loss function during the refocus model’s training phase. Nevertheless, it also acts as the first deep learning-based model that can synthesize realistic defocused images, which is useful in FQA, quality assurance and restoration applications.
By addressing these critical limitations, this thesis makes significant contributions to IQA and image restoration in the field of whole slide imaging. The introduction of the FQA and score fusion model not only boosts computational efficiency but also improves accuracy and generalizability. Additionally, the virtual refocusing model extends these improvements by tackling the defocus problem in WSI through precise adjustment of focus on a per-pixel basis.