| dc.description.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 dis-
tortions, 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 com-
plicate the restoration process for methods that lack an understanding of the 3D tissue
radiance. 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). Among the distortions that degrade
the quality of digital slides, out-of-focus blur is the most common one. Different from pho-
tographic images, WSIs have much bigger dimensions, making most deep-learning based
FQA models computationally infeasible. Based on prior knowledge of the WSI and its
imaging process, we developed a lightweight model named FocusLiteNN that is 10, 000
times more efficient than SOTA deep learning-based ones without compromising accu-
racy. Furthermore, we introduce the first open-source, expert annotated FQA dataset
TCGA@Focus, offering a comprehensive platform for developing and evaluating new FQA
models.
However, most FQA models, or IQA models in general, often exhibit biases towards
specific types of image content or distortions due to their different design principles or
training data. This poses a challenge for users when choosing the best quality assessment
model for their needs. A practical approach is to fuse the results of multiple existing IQA
models into a more robust one. Following this idea, we developed a novel framework for IQA core fusion that is able to select the best combination of models according to the
uncertainty in each image and the overall uncertainty of each model. This requires the
model to be equipped with both fine-grained uncertainty analysis at the content level and
coarse-grained uncertainty analysis at the model level, respectively. Existing models either
lack content-level uncertainty estimation or have limited generalizability due to supervised
training. Our method employs an unsupervised approach using deep Maximum a Posteriori
(MAP) estimation, which can be trained on a combination of multiple datasets without
the need for Mean Opinion Score (MOS). This greatly improves the generalizability of the
model.
The above two works address different problems in quality assessment. In practice,
detected bad-quality images are either rejected or recollected. In digital pathology, rec-
ollecting the biosample causes additional suffering for the patient. Consequently, defocus
restoration is a possible solution. Deblurring assumes that there exists a sharp image in
which all pixels are in-focus, which is commonly referred to as a All-In-Focus (AIF) image.
Although this assumption is true for natural images, it might not hold for WSIs due to its
transparency, uneven surface and the microscope’s shallow Depth of Field (DOF). Since
the target does not exist, WSI deblurring becomes an undefined task. We propose an
alternative approach to address the defocus problem, which is virtual refocusing. It aims
to simulate and surpass the traditional experience of one continuously adjusting the focus
of a microscope, allowing for a comprehensive examination of tissue structures at varying
depths without the need for physical slide presence. By implicitly learning a continuous
3D radiance representation from the sparse inputs, the proposed model can refocus each
pixel to any focus plane according to a focus map. As far as we know, this is the first work
on WSI virtual refocusing.
This thesis makes significant contributions to IQA and image restoration with applica-
tions in WSI. The introduction of the FocusLiteNN model boosts computational efficiency
while the score fusion model addresses the bias issue. 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. | en |
