Files
Abstract
Thyroid nodules are common and seen in over half the US population on autopsy. Though most thyroid nodules are benign, some do represent thyroid cancer and may be treated through thyroidectomy. Nodule risk-assessment begins first with radiologic evaluation on ultrasound. If indicated, nodule diagnosis continues with fine-needle aspiration (FNA) biopsy through which nodules can be grouped as benign, malignant, or indeterminate. Indeterminate thyroid nodules (ITNs), those neither conclusively benign nor malignant on FNA, make up roughly 30% of the nodules biopsied. Even with additional molecular testing performed, nearly half of all ITNs surgically excised are determined to have been benign on post-surgical pathology. There is a clinical need for additional diagnostic tools which aid in the classification of ITNs and reduce the number of ‘false-positive’ thyroid surgeries. This dissertation presents the development of artificial intelligence (AI) algorithms for the characterization and classification of indeterminate thyroid nodules on ultrasound. AI algorithms have the potential to segment and classify lesions on ultrasound; however, previous work in the field has focused primarily on FNA-determinate nodules. Therefore, the segmentation and classification of FNA-indeterminate nodules still present a large area of need within the current diagnosis paradigm. Further, as the role of thyroid nodule molecular testing platforms has expanded, there is an increased need for study into the integration of image-based AIs with molecular tests. This dissertation is separated into three primary studies conducted on a multi-institutional dataset. Segmentation AI for Thyroid Nodules on Ultrasound: An attention-enhanced U-Net algorithm was trained and independently tested for nodule segmentation performance on ultrasound. Additionally, a fuzzy c-means unsupervised algorithm was validated for internal segmentation of mixed thyroid nodules into solid and cystic components. Results of this study demonstrated high performance of the deep learning segmentation algorithm across all nodule pathologies. Classification AI for Indeterminate Thyroid Nodules on Ultrasound: Both radiomic and deep transfer learning classification algorithms for ITNs on US were developed. Multiple pre-training and feature selection approaches were investigated to improve algorithm performance. The impact of our automatic segmentation algorithms on radiomic algorithm performance was investigated. Finally, the integration of radiomic and deep learning transfer algorithms was explored for use in tandem with FNA biopsy and molecular testing. Results of this investigation showed a meaningful increase in surgical pathology classification performance when image-based AIs are merged with pre-surgical clinical pathology tests. Reference-Standard-Free Lesion Segmentation Performance Metric: The quality of nodule segmentation can impact feature calculation and downstream classification. Current segmentation performance metrics require the use of a reference-standard; however, this limits their use once an algorithm has been deployed. In this study, a reference-standard-free segmentation metric, lesion segmentation-sureness (LeSS) score, was developed and independently tested. LeSS was applied to AttU-Net of the first study and the radiomic model of the second study as a test case for its application. The results of this study showed high correlation between LeSS and traditional segmentation performance metrics. It also demonstrated that LeSS analysis could be used to improve segmentation performance and the calculation of two radiomic features.