| dc.description.abstract | Medical imaging provides a non-invasive way to visualize tissues and diseases, enabling
both qualitative and quantitative assessments that are needed in diagnosing and monitoring
a wide range of conditions. Modalities like computed tomography (CT), magnetic
resonance imaging (MRI), and ultrasound have become essential tools for capturing detailed
anatomical and functional information. In recent years, fully data-driven deep
learning approaches have shown significant potential in automated medical image analysis.
The key challenges included limited data availability, poor model generalizability,
long training times, and high latency for inference. Addressing these challenges is essential
to achieve quantitative improvements in tasks such as image reconstruction and
segmentation. This thesis work focuses on design and development of deep learning
models to enhance the quantitative performance of medical image analysis by addressing
these specific challenges. Two key applications that were considered in this thesis work
include (a). three-dimensional image reconstruction task of quantitative susceptibility
mapping (QSM) in MRI, and (b). median nerve segmentation in ultrasound imaging.
Quantitative Susceptibility Mapping (QSM) is an advanced MRI technique
that measures and visualizes the magnetic susceptibility of tissues, with key clinical
applications, especially in neuroimaging. It is widely used to visualize and quantify
iron deposits in the brain, which is crucial for studying neurodegenerative diseases like
Parkinson’s and Alzheimer’s. It also assesses myelin content in the white matter, which
is crucial for understanding conditions such as multiple sclerosis where demyelination
occurs. By providing both visualization and quantification, QSM is a valuable tool for
studying microstructural and compositional changes in the brain associated with various neurological conditions. These applications enhance diagnosis, treatment planning, and
monitoring of various medical conditions. Key step of QSM is reconstruction/estimate of
tissue magnetic susceptibility from a local field using the magnetic resonance (MR) phase
measurements. This requires solving the inverse problem between the measured magnetic
field distribution and the tissue magnetic susceptibility. This thesis work presents
three distinct works contributing to QSM reconstruction through novel model-based deep
learning frameworks, as well as pure deep learning approaches, each addressing unique
challenges of QSM reconstruction.
SpiNet-QSM: This part of the thesis work introduces a SpiNet-QSM framework, it
is a Schatten p-norm-driven model-based deep learning framework for QSM reconstruction
with a learnable norm parameter p and regularization parameter to adapt to the
data. It also has a 3D-CNN denoiser based trainable regularizer. In contrast to other
model-based architectures that enforce either l2-norm or l1-norm for the denoiser, the
proposed approach can enforce any p-norm (0 < p ≤ 2) on a trainable regularizer, which
adapts to the data. The proposed SpiNet-QSM showed a consistent improvement of at
least 5% in terms of the high frequency error norm (HFEN) and the normalized root
mean squared error (NRMSE) over other QSM reconstruction methods in limited training
data. Through evaluation on 93 imaging volumes with varied acquisition parameters,
SpiNet-QSM consistently demonstrated at least a 5% improvement in high-frequency
error norm (HFEN) and normalized root mean squared error (NRMSE) compared to
leading methods such as QSMnet and learned proximal convolutional neural network
(LPCNN), even under limited training data.
ISDU-QSMNet: This part of the thesis work introduces iteration specific denoising
via unshared weights for QSM Reconstruction, also referred to as ISDU-QSMNet,
an end-to-end model-based deep learning framework designed to effectively solve the
inverse problem of QSM reconstruction from the local field. ISDU-QSMNet introduces
significant modifications to existing model-based deep learning approaches by incorporating
unshared denoiser weights and random subset sampling during training, leading
to a more powerful, robust, and training-efficient model that improves the performance with full training data, reduces the overall training time, and effectively handles different
datasets. The proposed method was evaluated against other model-based deep learning,
as well as pure deep learning approaches, by performing reconstructions on 93 imaging
volumes with varying acquisition parameters under two scenarios: full training data and
limited training data. In the full training data scenario, the proposed approach demonstrated
substantial improvements over all existing methods in both model-based and
pure deep learning categories, achieving significant reductions in HFEN by upto 3.5%.
In the limited training data scenario, the proposed approach matched the performance of
state-of-the-art model-based deep learning models. Additionally, it demonstrated strong
generalization capabilities by effectively handling data with different acquisition parameters
and consistently performed well in ROI analysis.
MSFF-QSMNet:This part of the thesis work introduces MSFF-QSMNet, an endto-
end deep learning model based on a 3D-CNN architecture that combines a modified
two-level nested UNet structure (main network) with a multi-scale feature fusion module
(sub-network) to improve QSM reconstruction. The proposed MSFF-QSMNet was
evaluated against other deep learning-based approaches, including QSMnet, DeepQSM
and xQSM, as well as model-based deep learning approaches such as the LPCNN and
SpiNet-QSM by performing reconstructions on 93 imaging volumes with varying acquisition
parameters. This evaluation showed that MSFF-QSMNet is notably more effective
in minimizing both high-frequency errors and overall reconstruction errors, achieving
at least a 4% improvement in these key performance metrics compared to other stateof-
the-art QSM reconstruction methods, including both deep learning and model-based
approaches on full training data settings. MSFF-QSMNet demonstrates a clear advantage
over other deep learning methods in both full and limited data scenarios, showcasing
its effectiveness in handling QSM reconstruction tasks across both data availability conditions.
The median nerve is a major peripheral nerve that serves as a critical communication
pathway between the hand and central nervous system. This nerve plays a crucial
role in motor function. Additionally, it transmits sensory information, such as touch temperature, and pain, from the palm and fingers to the central nervous system. Given
its essential role in hand functionality, any compression of the median nerve can lead
to significant motor deficits and sensory disturbances. Carpal Tunnel Syndrome (CTS),
one of the most common conditions affecting the median nerve, is caused by increased
pressure within the carpal tunnel, compressing the nerve and leading to symptoms like
numbness, tingling, and weakness that impact daily activities. Segmenting the median
nerve in ultrasound imaging is essential in understanding and diagnosing CTS, as it allows
clinicians to measure the nerve’s cross-sectional area and shape within the carpal
tunnel, aiding in identifying nerve enlargement and compression points indicative of
CTS. This precise segmentation supports early diagnosis, tracks disease progression, and
evaluates treatment efficacy. Additionally, automated segmentation methods enhance diagnostic
efficiency, reduce reliance on invasive techniques like nerve conduction studies,
and improve patient comfort by providing a non-invasive and repeatable assessment.
MNSeg-Net: This part of the thesis study introduces MNSeg-Net, a novel lightweight,
2-level Nested UNet-based CNN model with 2.46 million parameters, which effectively
segments the median nerve in ultrasound frames, achieving average dice similarity coefficient
(DSC) scores of 94.7% at the wrist and 83.4% from wrist to elbow, and the
lowest Hausdorff distance, matching the performance of the best-performing 44 million
parameter heavy U2Net model. MNSeg-Net is a single deep learning model, specifically
designed to create a fully automated, end-to-end clinical setup that supports real-time
segmentation of the median nerve, including the computation of its cross-sectional area
(CSA). On a single GPU, MNSeg-Net can segment up to 42 frames in a second. This
work also involved development of a clinical setup using an Av.io HD Epiphan frame
grabber to stream ultrasound images to a GPU-based system, capable of processing up
to 32 frames per second on a single GPU. An additional screen, positioned alongside
the ultrasound monitor, displays processed frames with segmented median nerves and
their CSA, assisting clinicians in real-time diagnosis. This method enables real-time assessment
of the median nerve for CTS diagnosis, establishing a new standard for CTS
detection and management. Collectively, these studies resulted in deep learning models that enhance the quantitative performance in medical image analysis, specifically by
addressing common challenges such as limited training data, leveraging more training
data, the need for model generalizability, and real-time inference requirements. Although
these models have been designed for specific problems, the frameworks developed here
hold potential for much broader application across various medical imaging tasks. This
thesis work is a step toward advancing quantitative medical image analysis for practical
and reliable use in clinical settings. | en_US |
