Machine Learning Methods for the Analysis of Thyroid Nodule Ultrasound Images
Abstract
Thyroid nodules pose significant diagnostic challenges, particularly in pediatric populations and indeterminate cases. These challenges stem from variability in diagnostic practices, limited applicability of existing management guidelines for children, and the need for innovative, non-invasive tools to enhance risk stratification. This dissertation aims to address these limitations by leveraging machine learning (ML) techniques to improve the diagnosis and management of thyroid nodules using ultrasound imaging. The work focuses on four key areas: (1) investigating methods for training ML models on automatically curated ultrasound datasets, (2) validating management guidelines and ML algorithms for pediatric thyroid nodules, (3) assessing the diagnostic utility of cine imaging in thyroid ultrasound, and (4) enhancing risk stratification for indeterminate thyroid nodules through advanced ML techniques.
First, this study demonstrated the effectiveness of automatically curated datasets generated using the MADLaP tool, which combines natural language processing and computer vision to annotate large datasets. It was tested to have a 63\% yield rate and 83\% accuracy. However, the usefulness of the generated data for training deep learning models remains unknown. Hence, I conducted experiments to determine whether using an automatically curated dataset improves deep learning algorithms' performance. I trained deep learning models both on the manually annotated and on the automatically-curated datasets. I also trained with a smaller subset of the automatically-curated dataset that has higher accuracy to explore the optimum usage of such datasets. Deep learning algorithms trained on these datasets achieved a higher AUC (0.694) than those trained on smaller, manually curated datasets (AUC 0.643, p < 0.001) despite the inherent noise in automated annotations. The findings also showed that utilizing the full automatically curated dataset, rather than a smaller high-accuracy subset, maximized algorithmic performance. These results underscore the potential of automated curation to overcome data availability challenges and improve deep learning applications in thyroid nodule diagnosis.
Second, to address the unique needs of pediatric thyroid nodule management, this dissertation validated the performance of radiologists’ overall impressions, ACR TI-RADS, and a deep learning algorithm in differentiating benign and malignant nodules in a cohort of 139 children and young adults. The deep learning algorithm achieved a sensitivity of 87.5\%, outperforming radiologists (58.3\%) and ACR TI-RADS (85.1\%), though it had a lower specificity of (36.1\%). Both ACR TI-RADS and the deep learning algorithm had higher sensitivity and lower specificity than overall impressions. The study highlighted the potential role of deep learning as a screening tool for pediatric patients, where minimizing false negatives is critical due to the prolonged impact of untreated malignancies in young populations.
Third, the diagnostic value of cine imaging in thyroid ultrasound was systematically evaluated for the first time. Through a reader study involving 50 benign and 50 malignant nodules assessed by four specialty-trained radiologists, the addition of cine images to static images did not significantly impact the diagnostic performance, with sensitivity and specificity remaining consistent (static images: 0.65 and 0.20, static + cine: 0.67 and 0.22, p > 0.5). These findings suggest that current practice guidelines, which do not mandate cine imaging, are sufficient for accurate diagnosis and management. This result also implied that the development of AI tools could focus on static imaging to optimize computational resources and clinical workflows.
Lastly, the dissertation introduced a novel framework to enhance the diagnosis of indeterminate thyroid nodules, often classified as Bethesda III. They pose significant challenges due to their ambiguous imaging features and inconclusive biopsy results. By integrating soft labels and transfer learning into convolutional neural networks, the model achieved diagnostic performance comparable to molecular testing, offering a cost-effective and non-invasive alternative for risk stratification. This innovation has the potential to reduce unnecessary procedures while ensuring accurate identification of malignancies.
In conclusion, the outcomes of this dissertation contribute to the advancement of diagnostic and management strategies for thyroid nodules by addressing critical gaps in current methodologies. The findings have significant clinical implications, from reducing the burden of manual dataset annotation to improving pediatric and indeterminate nodule management. Furthermore, the methodologies developed here have broader applications in medical imaging, showcasing the transformative potential of artificial intelligence in healthcare.
Type
Department
Description
Provenance
Subjects
Citation
Permalink
Citation
Yang, Jichen (2024). Machine Learning Methods for the Analysis of Thyroid Nodule Ultrasound Images. Dissertation, Duke University. Retrieved from https://hdl.handle.net/10161/32637.
Collections
Except where otherwise noted, student scholarship that was shared on DukeSpace after 2009 is made available to the public under a Creative Commons Attribution / Non-commercial / No derivatives (CC-BY-NC-ND) license. All rights in student work shared on DukeSpace before 2009 remain with the author and/or their designee, whose permission may be required for reuse.