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

Abstract

Although reference ranges for 99m Tc 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.

Keeping Up With ChatGPT

“…In D (normal 18 F-DCFPyL PET), ChatGPT highlighted areas that show low uptake in truth (“bright”) incorrectly as “potential areas of increased uptake” (red circles added by ChatGPT). The images were modified from research that was originally published in J Nucl Med, as cited in 9,11,16,17 .…”

“…In D (normal 18 F-DCFPyL PET), ChatGPT highlighted areas that show low uptake in truth ("bright") incorrectly as "potential areas of increased uptake" (red circles added by ChatGPT). The images were modified from research that was originally published in J Nucl Med, as cited in 9,11,16,17 The number of cases with committing answers (ie, specific answer in relation to the image) and, among them, appropriate answers is shown for both versions of ChatGPT (percentage relative to the total of 15 cases). errors.…”

“…Fifteen images that are exemplary for common nuclear medicine examinations were obtained from different previous publications in the Journal of Nuclear Medicine [8][9][10][11][12][13][14][15] or the Journal of Nuclear Medicine 16,17 These images included examples of normal or abnormal findings for a scintigraphy (thyroid, bone, dopamine transporter, myocardial perfusion, and diuretic renal scintigraphy) and 4 examples of PET images ( 18 F-FDG, 18 F-DCFPyL). Images were cropped if needed.…”

Application of radiomics and machine learning to thyroid diseases in nuclear medicine: a systematic review

“…A total of 1643 articles were extrapolated with the computer literature search and, by reviewing the titles and abstracts, 1627 of them were excluded because the reported data were not within the field of interest of this review. Sixteen articles were therefore selected and retrieved in full-text version [18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33] and one additional study was also found screening the references of these articles (Fig. 1) [34]; as a consequence, the total number of studies evaluated in the review was 17.…”

The transformational potential of molecular radiomics

“…DL architecture to produce fourth‐order radiomic features from thyroid images. Image reprinted with permission Currie and Iqbal 2022 17 …”