Utilizing Multitask Transfer Learning for Sonographic Rheumatoid Arthritis Synovitis Grading

<p>Classifying the four sonographic Rheumatoid Arthritis (RA) synovitis grades (Grade 0, Grade 1, Grade 2, and Grade 3) is a difficult problem due to the complexity of the relevant markers. Therefore, the current research proposes a Multitask Transfer Learning (MTL) framework for sonographic RA synovitis grading of Ultrasound (US) images in Brightness mode (B-Mode) and Power Doppler mode.</p> <p>In the medical community, the lack of reliability of scoring these images has been an issue and reason for concern for doctors and other medical practitioners. The human/machine variability across the acquisition procedure of these US images creates an additional challenge that restricts the development of an efficient automated scoring system. The literature reports the lack of coherency among the doctors’ opinions about the grade of arthritis for patients in controlled trials. Motivated by these reasons, the current work moves away from the traditional wisdom of separately scoring B-mode and Power Doppler mode images and poses an MTL framework that jointly learns the features for images across both modes, leading to a more robust automated classifier. The multitask nature of the model also provides additional benefits such as better generalization to blinded test data from the inherent regularization and the ability of the model to be trained by a combination of B-Mode and Power Doppler US images, thereby efficiently handling data scarcity.</p> <p>Results show the superior performance of the proposed approach compared to traditional machine learning algorithms as well as other standard deep learning models such as Convolutional Neural Networks and Vision Transformers. The mean testing accuracy of our proposed MTL model on B-Mode was 51.55% and on Power Doppler was 61.18%. However, since the boundary between classes is not always clear or defined in RA synovitis grades, the Top 2 success rates have been regarded as another measure of performance in this domain. Accordingly, with a mean B-Mode Top 2 success rate of 80.52% and a Power Doppler Top 2 success rate of 82.50%, the proposed approach can reach a near-human doctor-level classification performance, establishing the usability of this approach.</p>

Name thesis:degree_name

Master of Science in Data Science

Level thesis:degree_level

Thesis - Open Access

Discipline thesis:degree_discipline

Mathematics

Year

2023

Author dc:creator

• Sanders, Jordan Marie Claire

• Medical Imaging
• Multitask Learning
• Deep Learning
• Arthritis Detection
• Transfer Learning
• Artificial Intelligence
• Data Science

Repository record dc:identifier

https://commons.erau.edu/edt/782

OAI identifier oai:identifier

oai:commons.erau.edu:edt-1811