BACKGROUND: Thyroid nodules are present in approximately 80% of adult patients. Ultrasound is the gold standard for initial assessment of thyroid nodules, with follow-up fine-needle aspiration (FNA) for nodules with suspicious sonographic features. Subsequent cytology is risk stratified by the Bethesda System. Bethesda III and IV are considered indeterminate and a standardized approach to management of patients with this cytology is lacking. Analysis for high-risk genes has emerged as an option to guide treatment in such cases. Our institution uses an internally validated 23-gene panel to identify all mutations in known hotspot regions by next-generation sequencing (NGS). Nodules with known high-risk mutations are referred for surgery, while others may be watched with ultrasound surveillance. As machine-learning continues to strengthen, there is a growing role for augmented medical imaging. Machine learning may aid in the early recognition of lesions likely to be genetically high risk by ultrasound alone.
PURPOSE: This study aims to evaluate whether an automated machine-learning algorithm can be used to retrospectively genetically risk-stratify thyroid nodules by ultrasound, using the presence of a high-risk mutation by NGS as the reference standard.
METHODS: Electronic medical records were obtained retrospectively from 105 patients who underwent ultrasound-guided FNA and NGS (using a 23-gene panel) for suspicious thyroid nodules from January 2017 through August 2018. Nodules were classified as high risk if a mutation was identified in a codon of known pathogenicity and low risk if no mutation was identified or if a mutation was identified in a region of unknown pathogenicity. High quality ultrasound images of the nodules were selected from the day of or within 6 months prior to the FNA with the assistance of a board-certified radiologist blinded to the NGS results; 508 ultrasound images across 101 lesions and 91 patients were extracted. 361 images were of low-risk genotype lesions and 147 images were of high-risk genotype lesions. Images were pre-processed and cropped in a blinded fashion. Machine learning was performed on Google AutoML beta software (Google Inc, Mountain View, CA). Images were randomly assigned to a training set (410, 81.7%), an internal validation set (48, 9.4%) and a final testing set (50, 9.8%).
RESULTS: Evaluative analysis of the model by Google AutoML was performed on the test set of 50 images. The model produced a sensitivity of 88.9% and specificity of 95.1% in the identification of genetically high-risk lesions by ultrasound. Only one lesion that was high risk was misclassified as low-risk, and just two lesions that were low risk were misclassified as high risk. Of note, surgical pathology on the one misclassified high-risk lesion was follicular adenoma with oncocytic features. Positive and negative predictive values were 80% and 95%, respectively.
CONCLUSION: An automated machine learning classifier possesses the ability to discriminate sonographic images of thyroid nodules based on genotype risk. These preliminary results advocate for further exploration with greater control over technical and patient variables. Broad future directions include the potential for augmented image interpretation and feature extraction to identify additional high risk sonographic characteristics of thyroid nodules.