Remodeling 99mTc-Pertechnetate Thyroid Uptake: Statistical, Machine Learning, and Deep Learning Approaches

Although reference ranges for Tc-99m thyroid percentage uptake vary, the seemingly intuitive evaluation of thyroid function does not reflect the complexity of thyroid pathology and biochemical status. The emergence of artificial intelligence in nuclear medicine has driven problem solving associated with logic and reasoning, warranting reexamination of established benchmarks in thyroid functional assessment. Methods: This retrospective study of 123 patients compared scintigraphic findings with grounded truth established through biochemistry status. Conventional statistical approaches were used in conjunction with an artificial neural network to determine predictors of thyroid function from data features. A convolutional neural network was also used to extract features from the input tensor (images). Results: Analysis was confounded by subclinical hyperthyroidism, primary hypothyroidism, subclinical hypothyroidism, and triiodothyronine toxicosis. Binary accuracy for identifying hyperthyroidism was highest for thyroid uptake classification using a threshold of 4.5% (82.6%), followed by pooled physician interpretation with the aid of uptake values (82.3%). Visual evaluation without quantitative values reduced accuracy to 61.0% for pooled physician determinations and 61.4% classifying on the basis of thyroid gland intensity relative to salivary glands. The machine learning (ML) algorithm produced 84.6% accuracy; however, this included biochemistry features not available to the semantic analysis. The deep learning (DL) algorithm had an accuracy of 80.5% based on image inputs alone. Conclusion: Thyroid scintigraphy is useful in identifying hyperthyroid patients suitable for radioiodine therapy when using an appropriately validated cutoff for the patient population (4.5% in this population). ML artificial neural network algorithms can be developed to improve accuracy as second-reader systems when biochemistry results are available. DL convolutional neural network algorithms can be developed to improve accuracy in the absence of biochemistry results. ML and DL do not displace the role of the physician in thyroid scintigraphy but can be used as second-reader systems to minimize errors and increase confidence.

Automated summary

This study evaluated thyroid function by comparing scintigraphic findings with biochemistry status. Results showed that thyroid uptake classification and physician interpretation with the aid of uptake values were effective in identifying hyperthyroid patients. Machine learning and deep learning algorithms can improve diagnostic accuracy, especially in the absence of biochemistry results.