TY  - BOOK
AU  - Peter Emad Salah,
AU  - Inas Ahmed Yassine 
TI  - Noninvasive detection and quantification of liver fibrosis from strain-encoded elastography using artificial neural network
U1  - 610.28 
PY  - 2024///
KW  - Biomedical Engineering
KW  - qrmak
KW  - liver fibrosis
KW  - magnetic resonance elastography
KW  - liver stiffness measurement
KW  - elastogram
KW  - artificial neural networks
N1  - Thesis (M.Sc.)-Cairo University, 2023.; Bibliography: pages 49-55.; Issued also as CD
N2  - Liver fibrosis is a common reaction to very different diseases and infection of liver
tissue. Several factors have caused liver fibrosis to spread all over the world. In some
countries, food and water contaminants have caused large amounts of liver infections [1].
In other areas of the world wrong diet habits that cause obesity or increase the amount of
alcohol consumption can lead to liver fibrosis too [2]. Liver diseases are very common
especially in third world countries where hepatitis B and hepatitis C only are widely
spread affecting more that 429 million patients while the percentage of diagnosed people
does not exceed 10% [3].
In particular, Chronic Liver Diseases (CLD) are estimated to affect 1.5 billion people
worldwide. Progression of CLD leads to excessive fibrosis and cirrhosis which could
lead to hepatocellular carcinoma. Accurate diagnosis and staging of fibrosis are essential
for proper disease prognosis and progression monitoring [4]. Liver biopsy is considered
the gold standard for grading and staging of fibrosis [5]. However, it is an invasive and
expensive procedure that could cause complications, pain for the patient, and suffers from
sampling errors and observer variability.
Alternatively, Magnetic Resonance Elastography (MRE) is considered the best
noninvasive technique for LSM and fibrosis grading [6]. MRE produces strain-encoded
images indicating the Liver Stiffness Measurement (LSM) distribution in the liver and
they significantly correlate with biopsy results without suffering from observer
variability. However, MRE staging is limited by different cut-off thresholds in different
studies and requires radiologists to manually select a valid ROI from the liver for LSM
estimation. Consequently, a pipeline is presented to automatically segment the liver from
an MRE scan, select a high-quality Region of Interest (ROI) and estimate LSM without
any radiologist intervention. First, a U-Net was used to segment the liver from the MRE
magnitude images. Then, the MRE Non-linearity images were used as confidence map
indicators to determine the high SNR regions. Then they were thresholded using Otsu
algorithm and then inverted to mark high quality regions. Additionally, in order to extract
a high-quality ROI, the Non-linearity output and the liver masks were combined together.
Finally, the high-quality ROIs from Gelastic, Velocity, Viscosity and Elasticity MRE
strain-encoded images were used to calculate the LSM using weighted averaging. The
extracted measurements were used to feed an Artificial Neural Networks (ANN) for the
automatic grading of liver fibrosis, with liver biopsy grades as the true labels.
The LSM estimation reached correlations of 0.92, 0.89, 0.90, 0.81 with the
radiologists’ estimation using the MRE Gelastic, viscosity, elasticity, and velocity
images respectively. The ANN was able to achieve accuracy of 85% for fibrosis grading
using conventional train, validation, test data splitting. Finally, a data sampling approach
is used to increase the amount of data with high-quality samples. The proposed technique
allows further experiments and investigations. It has been then used along with neural
network search to compare between different approaches that aim to find the optimum
system that is capable of providing an accurate and robust system for liver fibrosis
grading. It was found that patch-based training with CNN can provide the most robust
system and the combination of Elasticity, Velocity and Viscosity can provide the best
combination toward highest accuracy; الأمراض المزمنة للكبد تؤثر على 1.5 مليار شخص حول العالم وتؤدي إلى تليّف وسرطان الكبد، لذلك التشخيص الصحيح مطلوب. ولذلك، يقدم هذا البحث حلا لتجزئة الكبد تلقائيًا، واختيار منطقة ذات جودة عالية من الاهتمام، وتقدير معامل مرونة الكبد بدون أي تدخل من أخصائي الأشعة. يتم استخراج معلومات مرونة الكبد من صور الاشعة ومن ثم إرسالها إلى شبكة عصبية اصطناعية لتقدير تصنيف التليف الكبدي تلقائيًا. وحققت الشبكة العصبية الاصطناعية دقة بلغت 85% في تصنيف التليف الكبدي. وأخيرًا، يتم استخدام تقنية لزيادة كمية العينات ذات الجودة العالية فتمكنا من إجراء تجارب و استقصاءات إضافية للعثور على النظام الأمثل والأوفر دقيقًة لتصنيف التليف الكبدي.
ER  -