Remedies for challenges in deep learning-based quantitative ultrasound imaging

The thesis of this dissertation suggests that there is a continuous demand for the development of data-efficient and robust deep learning (DL) algorithms in order to promote wider adoption of DL-based algorithms in clinical imaging. This dissertation proposes solutions specifically to address challenges in Deep Learning-based Quantitative Ultrasound (QUS). QUS is a technique that uses ultrasound to estimate the properties of tissues, such as their scattering properties, elasticity or other relevant parameters, for the purposes of classifying tissue state. DL-based QUS aims to improve the accuracy and efficiency of this technique by using deep neural networks to analyze ultrasound image data and extract quantitative information. DL-based QUS approaches have the advantage of being model free and calibration-free if the data is acquired under a scenario of constant image settings from a single machine. However, DL-based QUS faces several challenges, including the need for large amounts of data to train deep neural networks, the variability of ultrasound images due to differences in equipment and imaging settings, and the need for enhancing the understanding of security aspect of DL algorithms. To address these challenges, this dissertation proposes the development of a data-efficient deep learning algorithm called Zone Training, which aims to learn diffraction patterns separately. Additionally, the study introduces a Transfer Function to mitigate variability between ultrasound images resulting from differences in equipment and imaging settings. Building on this, a strategy is developed to "steal" functionality from a victim machine and implement it on a perpetrator machine. Overall, this study highlights the importance of developing data-efficient and robust DL algorithms for the wider adoption of DL-based quantitative ultrasound (QUS). It proposes several remedies to address the challenges faced by this emerging field. Additionally, it demonstrates the ease of stealing the functionality of deep learning models, underscoring the need for security development of these models in clinical settings.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Soylu, Ufuk

Contributors dc:contributor

• Oelze, Michael L.
• Do, Minh N.
• Song, Pengfei
• Chen, Yun-Sheng
• Chandrasekaran, Varun

• Deep Learning
• Quantitative Ultrasound Imaging
• Tissue Characterization
• Data Scarcity
• Data Mismatch
• Model Security

Statement dc:rights

• Copyright 2024 Ufuk Soylu

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/124377

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F124377