Combined genetic and ultrasound image-based classifier for thyroid cancer diagnosis

A combined genetic and ultrasound image-based classifier addresses the issue of overtreatment in thyroid cancer diagnosis by accurately classifying nodules, reducing unnecessary procedures and costs through polygenic risk scores and deep learning analysis.

WO2026236419A1PCT designated stage Publication Date: 2026-11-12THE REGENTS OF THE UNIVERSITY OF COLORADO
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/026989
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-10-07
Filing Date
2026-05-07
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Current diagnostic frameworks for thyroid cancer, such as TI-RADS and FNA biopsies, result in high rates of overtreatment, with 72% of biopsies being benign and 20% indeterminate, leading to unnecessary procedures and significant healthcare costs.

Method used

A combined genetic and ultrasound image-based classifier that calculates hereditary predisposition to thyroid cancer using polygenic risk scores, integrating genetic assessment with deep learning analysis of thyroid nodule ultrasound features to accurately classify nodules as benign or malignant, reducing the need for unnecessary biopsies.

Benefits of technology

This approach reduces the number of unnecessary FNA procedures by 175,000 to 240,000 annually, preventing inadequate and indeterminate cytology diagnoses and diagnostic thyroidectomies, thereby lowering healthcare costs and improving clinical decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026026989_12112026_PF_FP_ABST
    Figure US2026026989_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A computer-implemented method and system classify a thyroid nodule using genetic and ultrasound image data. Genetic variant data are processed to determine a first polygenic risk score associated with thyroid cancer and a second polygenic risk score associated with benign thyroid nodular disease. A composite risk metric is generated from the first and second polygenic risk scores to reflect differential genetic contributions to malignant and benign thyroid nodular conditions. Ultrasound image data of the thyroid nodule may be processed using a trained deep learning model to generate an image-based risk assessment, including classification as benign, papillary-thyroid-cancer-like, or follicular-thyroid-cancer-like. The composite risk metric and image-based risk assessment may be combined to generate a classification or recommendation regarding biopsy, observation, surveillance, or intervention. The polygenic risk score may identify subjects at risk for aggressive thyroid cancer.
Need to check novelty before this filing date? Find Prior Art

Description

COMBINED GENETIC AND ULTRASOUND IMAGE-BASED CLASSIFIER FOR THYROID CANCER DIAGNOSISCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Patent Application No.63 / 801,737, titled COMBINED GENETIC AND ULTRASOUND IMAGE-BASED CLASSIFIER FOR THYROID CANCER DIAGNOSIS, filed May 7, 2025, and claims priority to U.S. Provisional Patent Application No. 63 / 894,950, titled COMBINED GENETIC AND ULTRASOUND IMAGE-BASED CLASSIFIER FOR THYROID CANCER DIAGNOSIS, filed October 7, 2025, both of which are hereby incorporated by reference in its entirety.FIELD

[0002] The present inventive concept relates to classifiers for use in and methods of diagnosis of thyroid cancer.STATEMENT REGARDING FEDERAL SUPPORT

[0003] This invention was made in part with government support under grant number R21-CA282380 awarded by the National Cancer Institute. The government has certain rights in the invention.BACKGROUND

[0004] Thyroid cancer is the most common endocrine malignancy, with -44,000 new cases representing 1% of new cancer diagnoses annually in the United States. Clinical evaluation for thyroid cancer starts when a thyroid nodule is suspected on physical examination or found incidentally on medical imaging. Thyroid ultrasound is performed to assess thyroid nodule sonographic characteristics, and clinical risk stratification algorithms such as the American College of Radiology Thyroid Imaging Reporting and Data System (TI-RADS) or the American Thyroid Association risk patterns are applied to decide on the need for the fine needle aspiration (FNA) biopsy to rule out cancer. More than 600,000 FNAs are performed in the United States annually at an estimated cost of $480 - $720 million.

[0005] Despite the wide adoption of TI-RADS, only 8% of biopsy results fit cytologic categories that require thyroid surgery; 72% produce benign results, and the remaining 20% are either inadequate or indeterminate, necessitating repeat procedures, expensive molecular testing, and / or diagnostic surgery. More than half of diagnostic thyroidectomies performed in patients with indeterminate thyroid nodules result in benign pathology.

[0006] Accordingly, there remains a need for improved frameworks that reduce overtreatment of patients with thyroid nodules and result in major savings for the healthcare system, resulting in a benefit to patients, physicians, and payers.SUMMARY

[0007] Aspects and implementations of the inventive concept relate to a framework that can estimate the risk of thyroid cancer in patients with thyroid nodules before an FNA is performed. This is achieved by calculating hereditary predisposition to developing thyroid cancer, which contributes up to 53% to the risk of disease. The genetic assessment is combined with deep learning analysis of thyroid nodule ultrasound features. The two approaches are complementary and synergistic. These risk assessments allow confidently excluding malignancy in a substantial proportion (e.g., -50%) of benign thyroid nodules. In some implementations, the polygenic risk score is further used to identify subjects at elevated risk for aggressive thyroid cancer, including risk associated with structural disease recurrence, tumor size, tumor focality, lymph node metastasis, extranodal extension, locoregional metastasis, or death from thyroid cancer. The polygenic risk score can be used to augment clinical ultrasound-based risk stratification schemas, such as TI-RADS and the American Thyroid Association algorithm, to determine which thyroid nodules are more likely to be malignant and require fine needle aspiration biopsy. A polygenic risk score can be used for molecular diagnostic testing of thyroid nodules with an indeterminate cytologic diagnosis to improve clinical decision-making of whether to proceed with diagnostic thyroidectomy or continue active surveillance.

[0008] The implementation of such a system in the United States alone can reduce the number of FNA for benign thyroid nodules by about 175,000 to about 240,000, depending on the selected sensitivity / specificity threshold and clinical implementation assumptions, and considering a total of 600,000 FNA procedures per year and a 20% malignancy rate.Consequently, it can prevent many inadequate and indeterminate cytology diagnoses, molecular diagnostic tests and diagnostic thyroidectomies for benign thyroid nodules.

[0009] In summary, the aspects and implementations of the inventive concept significantly reduce overtreatment in patients with thyroid nodules and result in major benefits and savings for the healthcare system.BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Non-limiting and non-exhaustive examples are described with reference to the following figures. Throughout the drawings, reference numbers can be re-used to indicate correspondence between referenced elements. The drawings are provided to illustrate embodiments of the present disclosure and do not limit the scope thereof.

[0011] FIG. 1 illustrates pleiotropic and phenotype-specific loci associated with thyroid diseases in a meta-analysis of genome-wide association studies (GWAS), in accordance with some aspects of the present disclosure.

[0012] FIG. 2 illustrates a graph showing mRNA expression of thyroid cancer-associated genes in normal thyroid tissue and thyroid cancer tissue, in accordance with some aspects of the present disclosure.

[0013] FIG. 3 illustrates graphs and a table showing data related to thyroid cancer polygenic risk scores, in accordance with some aspects of the present disclosure.

[0014] FIG. 4 illustrates data related to germline genetic susceptibility to thyroid cancer and benign nodular goiter, in accordance with some aspects of the present disclosure.

[0015] FIG. 5 illustrates a graph showing a scatterplot of effect sizes of variants in PTCSC2 locus significantly associated with thyroid cancer and benign nodular goiter, in accordance with some aspects of the present disclosure.

[0016] FIG. 6 illustrates a graph showing thyroid cancer polygenic risk score performance by ancestry, in accordance with some aspects of the present disclosure.

[0017] FIG. 7 illustrates a flowchart depicting a method for classifying a thyroid nodule in a subject, in accordance with some aspects of the present disclosure.

[0018] FIG. 8 illustrates a flowchart depicting a method for classifying a thyroid nodule using combined genetic and ultrasound image-based assessment, in accordance with some aspects of the present disclosure.

[0019] FIG. 9 illustrates a flowchart depicting a method for estimating a risk of aggressive thyroid cancer for a subject based on a polygenic risk score, in accordance with some aspects of the present disclosure.

[0020] FIG. 10 illustrates a flowchart depicting a method for estimating a risk of aggressive thyroid cancer for a subject based on a polygenic risk score with threshold-based classification, in accordance with some aspects of the present disclosure.

[0021] FIG. 11 illustrates a flowchart depicting a method for classifying a thyroid nodule using combined genetic and ultrasound image-based assessment, in accordance with some aspects of the present disclosure.

[0022] FIG. 12 illustrates a flowchart depicting a method for post-GWAS quality control processing, in accordance with some aspects of the present disclosure.

[0023] FIG. 13 illustrates a block diagram of a system for classifying a thyroid nodule, in accordance with some aspects of the present disclosure.

[0024] FIG. 14 illustrates a block diagram of a computer model system depicting computer model inference and training operations, in accordance with some aspects of the present disclosure.

[0025] FIG. 15 illustrates a block diagram of a neural network architecture, in accordance with some aspects of the present disclosure.

[0026] FIG. 16 illustrates a block diagram of a deep learning model classification architecture for thyroid nodule assessment, in accordance with some aspects of the present disclosure.DETAILED DESCRIPTION

[0027] Although certain embodiments and examples are described below, it will be understood that the disclosure extends beyond the specifically disclosed embodiments and / or uses and obvious modifications and equivalents thereof. Thus, it is intended that the scope of the disclosure herein disclosed should not be limited by any particular embodiments described below. It should be recognized that such description is not intended as a limitation on the scope of the present disclosure. Rather, the description also encompasses combinations and modifications to the embodiments and the examples described below.

[0028] As used herein, the term "composite risk metric" refers to a numerical value derived from a combination of two or more polygenic risk scores that reflects the differential genetic contributions to distinct disease conditions. The composite risk metric may be computed as a difference, a weighted combination, a ratio, or any other mathematical operation that combines the individual polygenic risk scores. For example, the composite risk metric may be computed as the difference between a first polygenic risk score associated with thyroid cancer and a second polygenic risk score associated with benign thyroid nodular disease. In other examples, the composite risk metric may be computed as a weighted sum of the first and second polygenic risk scores, where the weights may be determined empirically or through optimization.

[0029] As used herein, the term "differential genetic contributions" refers to the distinct sets of genetic variants and biological pathways that contribute to different disease phenotypes. For example, certain genetic variants may contribute predominantly to thyroid cancer risk while other genetic variants may contribute predominantly to benign thyroid nodular disease risk, and still other genetic variants may contribute to both conditions. The differential genetic contributions may be identified through genome-wide association studies, transcriptome-wide association studies, pathway analyses, or other genetic analyses.

[0030] As used herein, the term "clinical decision-making" refers to the process by which a healthcare provider determines a course of action for a patient based on available clinical information. Clinical decision-making may include, but is not limited to, determining whether to perform a fine needle aspiration biopsy, recommending observation or active surveillance, recommending surgical intervention such as lobectomy or thyroidectomy, ordering molecular diagnostic testing, or adjusting postoperative management.

[0031] As used herein, the term "image-based risk assessment" refers to a quantitative or qualitative evaluation of the likelihood that a thyroid nodule is malignant, generated by processing ultrasound image data through a computational model such as a trained deep learning model. The image-based risk assessment may comprise a probability score, a classification label, a risk category, or any other output indicative of the malignancy risk of the thyroid nodule.

[0032] As used herein, the term "combined classification" or "combined risk estimate" refers to an integrated assessment that incorporates both genetic risk information and imagebased risk information to classify a thyroid nodule or estimate a subject's risk. The combinedclassification may be generated by a m eta-classifier that receives as inputs the composite risk metric and the image-based risk assessment and produces a unified output.

[0033] As used herein, the term "aggressive thyroid cancer" refers to thyroid cancer that exhibits one or more high-risk clinical features, including but not limited to large tumor size, multifocal disease, lymph node metastasis, extranodal extension, extrathyroidal extension, distant metastasis, vascular invasion, structural disease recurrence, or death from thyroid cancer. Aggressive thyroid cancer may include, but is not limited to, thyroid cancers classified as intermediate-risk or high-risk according to the American Thyroid Association risk stratification system.

[0034] As used herein, the term "predefined threshold" refers to a cutoff value applied to a polygenic risk score to classify a subject into a risk category. The predefined threshold may correspond to a specific percentile or decile of the polygenic risk score distribution in a reference population, such as the top decile. In other examples, the predefined threshold may be determined based on clinical considerations, such as a desired sensitivity or specificity for identifying subjects at elevated risk.

[0035] As used herein, the term "high risk" when used in reference to a subject's classification based on a polygenic risk score, refers to a determination that the subject's polygenic risk score exceeds a predefined threshold, indicating an elevated genetic predisposition to developing aggressive thyroid cancer relative to the reference population.

[0036] As used herein, the term "calibrated" when used in reference to a polygenic risk score, refers to the adjustment of the polygenic risk score to account for differences in genetic ancestry, population structure, or demographic characteristics among subjects. Calibration may involve rescaling the mean and variance of the polygenic risk score within control participants using an adjustment model that integrates genetic components of ancestry. In some examples, calibration may involve fitting a regression model that accounts for genetic principal components, age, sex, or other covariates to produce an ancestry-corrected polygenic risk score.

[0037] FIG. 1 illustrates pleiotropic and phenotype-specific loci associated with thyroid diseases in a meta-analysis of genome-wide association studies (GWAS), in accordance with some aspects of the present disclosure. More specifically, the heatmap 100 illustrates the genetic correlation (r2) between thyroid phenotypes, which was estimated using covariate-adjusted LD score regression. Circular plots 102 and 104 highlight loci significantly associated with thyroidcancer and benign nodular goiter (right), and autoimmune thyroid diseases (left). Pleiotropic loci are marked by dots. In the circular plot 104, dots and gene labels show loci predominantly associated with thyroid cancer or benign nodular goiter, respectively. In the circular plot 102, dots show loci significantly associated with Graves’ disease, but not lymphocytic thyroiditis or primary hypothyroidism. As shown in the circular plot 104, the PTCSC2 is the only locus inversely associated with thyroid cancer and benign nodular goiter.

[0038] In some examples, the SNP-based heritability, estimated using covariate-adjusted linkage disequilibrium score regression, may range from approximately 0.07 for benign nodular goiter to approximately 0.11 for hypothyroidism in mixed-ancestry meta-analyses. These heritability estimates can help quantify the proportion of phenotypic variance attributable to common genetic variants and provide a foundation for the polygenic risk score approach.

[0039] In some examples, certain loci may be discovered in thyroid cancer meta-analysis but not in benign nodular goiter GWAS. These thyroid cancer-specific loci, which include ATM, CHEK2, TP53 and others, are associated with malignant transformation and allow discriminating between benign and malignant thyroid nodules in accordance with some aspects of the present disclosure

[0040] FIG. 2 illustrates a graph 200 showing mRNA expression of thyroid cancer-associated genes in normal thyroid tissue and thyroid cancer tissue, in accordance with some aspects of the present disclosure. Genes were identified from ANNOVAR annotations of genome-wide significant variants in thyroid cancer GWAS meta-analysis and FUSION TWAS cis-eQTL analysis. Indicator 202 illustrates that the gene associated with indicator 202 has a high expression in normal thyroid tissue (thyroid is among the top three highest expressing tissues in pan-tissue transcriptome analysis from the NCBI Gene database https: / / www.ncbi.nlm.nih.gov / gene). Significant associations (regression analysis adjusted for major somatic oncogenic driver [BRAF V600E vs. N / H / KRAS\ at Bonferroni-corrected p-value < le-04; indicator 204 illustrates positive; indicator 206 illustrates negative) of mRNA expression with high-risk thyroid cancer features such as earlier age at diagnosis, higher ERK score and lower thyroid differentiation score.

[0041] In some examples, a transcriptome-wide association study (TWAS) can be performed using cis-acting expression quantitative trait locus (cis-eQTL) analysis to identify potential causal variants affecting gene expression and to assign intergenic and noncoding RNAvariants to protein-coding genes. In an illustrative example, the TWAS may be performed using FUSION and Summary -based PrediXcan (S-PrediXcan) methods with thyroid tissue expression models from the Genotype-Tissue Expression (GTEx) project. The TWAS may identify the expression of protein-coding genes significantly associated with thyroid cancer, including genes not reaching genome-wide significance in the GWAS meta-analysis alone. For example, the TWAS may identify VEGFC an NBRl as additional genes associated with thyroid cancer risk, further expanding the set of genes that may be used in polygenic risk score calculations. In some examples, the TWAS may attribute noncoding RNA intronic variants to the expression of protein-coding genes such as TGFB2, which may be associated with all thyroid diseases analyzed. Many genes discovered in the TWAS may be found in more than one thyroid phenotype, consistent with the genetic overlap between thyroid diseases.

[0042] In some examples, the TWAS can identify a substantial number of protein-coding genes associated with thyroid cancer. In an illustrative example, the FUSION TWAS, as applied to mixed-ancestry and European-like thyroid cancer GWAS meta-analysis data, can identify the expression of approximately 55 unique protein-coding genes. The TWAS can also identify approximately 47 and 45 significant (after Bonferroni adjustment) lead cis-eQTL variants from the mixed-ancestry and European-like GWAS, respectively. For example, the TWAS attributed many significant intergenic and noncoding variates to protein-coding genes based on reported eQTL stats. For example, noncoding RNA intronic variant 1 :218515813 :T:C (mixed-ancestry ThC GWAS meta-analysis P=4.07 X 1 O'39) was attributed to the expression of TGFB2 in the TWAS (P=3.597X10‘61). Most significant genes found by the FUSION TWAS were also be replicated by S-PrediXcan, indicating analytical rigor. These TWAS findings can expand the set of genes that may be used in polygenic risk score calculations and provide additional biological context for the genetic architecture of thyroid cancer.

[0043] In some examples, pathway analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome databases may be performed on genes identified from the GWAS meta-analysis and the TWAS. The pathway analysis can help identify cell cycle, senescence and apoptosis as key biological processes contributing to thyroid cancer risk. The IGF1 and PI3K / Akt signaling pathways may be significantly associated with benign nodular goiter. These pathway-level findings provide additional biological context for the differentialgenetic architecture between thyroid cancer and benign nodular goiter that underlies the composite risk metric approach.

[0044] FIG. 3 illustrates graphs 302-306 and a table 308 showing data related to thyroid cancer polygenic risk scores, in accordance with some aspects of the present disclosure. In an illustrative example, two thyroid cancer polygenic risk scores (PRS) were developed: PRSThC vs. All to identify individuals at risk in a population and PRSThC vs. BNG for the clinically relevant task of discriminating malignant and benign thyroid nodules. PRS were tested in the Colorado Center for Personalized Medicine Biobank (CCPM) population (n = 94,651), and CCPM data was not used for PRS development. PRSThC vs. All was derived using the PRS-CS method. PRSThC vs. BNG was calculated using independent significant associations with thyroid cancer and benign nodular goiter. Graphs 302 and 304 illustrate receiver operating characteristic curves. Graph 306 illustrates thyroid cancer risk by PRS decile. Table 308 illustrates PRS association with features of aggressive thyroid cancer. * - p-value < 0.05. ** - p-value < 1.7e-3 (Bonferroni-corrected significance threshold).

[0045] In some examples, the adjusted PRS incorporating covariates of age, sex and genetic principal components may be cross-validated using five-fold cross-validation to ensure robustness and minimize overfitting. The PRS performance predicting binary phenotypes may be assessed using the area under the receiver operating characteristic curve, and AUCs may be compared using DeLong's test for significant differences.

[0046] In some examples, incorporating demographic and genetic ancestry covariates may improve PRS predictions. More specifically, incorporating age, sex and genetic ancestry covariates may improve the AUC for thyroid cancer identification to approximately 0.72, because, in part, of the higher incidence of thyroid disease in women and the increased risk of developing thyroid nodules and hypothyroidism with age. However, in some other examples, no significant improvement in clinical PRS performance for distinguishing thyroid cancer from benign nodular goiter may be observed when demographic covariates are added.

[0047] In some examples, incorporating demographic and genetic ancestry covariates may improve the AUC for thyroid cancer identification from approximately 0.69 to approximately 0.725 (95% CI 0.708-0.742). For other thyroid diseases, the covariate-adjusted AUC may range from approximately 0.690 (95% CI 0.662-0.718) for benign nodular goiter to approximately 0.729 (95% CI 0.714-0.745) for hypothyroidism.

[0048] FIG. 4 illustrates data 400 related to the germline genetic susceptibility of a person to thyroid cancer and benign nodular goiter, in accordance with some aspects of the present disclosure. Two biological processes with distinct genetic architecture may cause thyroid nodules: 1) hyperplasia, a polyclonal follicular cell proliferation with no malignant potential, and 2) neoplasia, a clonal growth driven by somatic genetic alterations. Neoplastic nodules can be benign or malignant, and the mismatch between biological mechanisms (hyperplasia vs. neoplasia) and GWAS phenotype definitions (benign and malignant thyroid nodules) caused apparent genetic pleiotropy. Pathway and genes associated with benign nodular goiter but not thyroid cancer in GWAS meta-analysis (e.g., insulin-like growth factor 1 [IGF1] and fibroblast growth factor [FGF] signaling pathways) predispose to benign nodules (green background). Pathways and genes associated with both benign nodular goiter and thyroid cancer (e.g., telomere maintenance) predispose to neoplastic thyroid nodules, benign or malignant (orange background). In the absence of other genetic risk factors, patients develop benign adenomas or low-risk thyroid cancers. Alternatively, genetic alterations in cell cycle and DNA damage response genes (associated predominantly with thyroid cancer but not benign nodular goiter in GWAS meta-analysis) can predispose a person to high-risk thyroid cancer.

[0049] In some examples, among the significant exonic variants identified in the meta-analysis, a majority may be nonsynonymous, potentially altering protein function. For example, among 46 significant exonic variants, 43 may be nonsynonymous. These nonsynonymous variants may directly affect the structure or function of the encoded proteins, providing mechanistic support for their role in thyroid disease susceptibility.

[0050] In some examples, ancestry-stratified GWAS meta-analyses may identify additional rare variants associated with thyroid cancer. More specifically, a rare nonsynonymous exonic variant in the shelterin complex gene TERFI (MAF = 0.0007) may be significantly associated with thyroid cancer in European-like populations. Such rare variants in telomere-associated genes may provide additional support for the role of telomere maintenance in thyroid cancer susceptibility.

[0051] In some examples, ancestry-stratified analyses may also identify variants in genes with known roles in thyroid hormone metabolism. More specifically, a variant in DIO1, which encodes an enzyme that converts pro-hormone thyroxine to the active thyroid hormone triiodothyronine, may be significantly associated with hypothyroidism in European-likepopulations. Such mechanistically plausible associations support the biological relevance of the genetic variants identified in the meta-analysis.

[0052] In some examples, additional thyroid cancer (ThC)-specific loci may include variants in the cell cycle regulator gene CDC25B. The identification of multiple cell cycle-related genes among ThC-specific loci supports the importance of cell cycle regulation in the malignant transformation of thyroid follicular cells. Further, in some examples, loci such nsHAUS6, which is involved in microtubule attachment to the kinetochore and central spindle formation, and SDCCAG8, which encodes a centrosome-associated protein, may demonstrate a stronger association with thyroid cancer despite the greater statistical power of the benign nodular goiter GWAS meta-analysis, further supporting the differential genetic architecture between thyroid cancer and benign nodular goiter.

[0053] FIG. 5 illustrates a graph 500 showing a scatterplot of effect sizes of variants in the PTCSC2 locus significantly associated with thyroid cancer and benign nodular goiter, in accordance with some aspects of the present disclosure. As illustrated in FIG. 5, benign nodular goiter significant associations effect size decrease as the thyroid cancer significant associations effect size increases. The Spearman correlation can be about -0.769.

[0054] FIG. 6 illustrates a graph 600 showing thyroid cancer polygenic risk score performance by ancestry, in accordance with some aspects of the present disclosure. As illustrated in FIG. 6, the area under the receiver operating characteristic curve (AUC) and 95% confidence interval are shown for PRSThC vs. BNG developed from the leave-CCPM biobank-out thyroid cancer meta-analysis and tested on CCPM biobank participants of various ancestries. Case and control counts are shown within bars. * - p-value < 0.05; ** - p < 0.01; *** - p-value < 0.001.

[0055] FIG. 7 illustrates a flowchart depicting flow 700 for classifying a thyroid nodule in a subject, in accordance with some aspects of the present disclosure. At 702, genetic variant data associated with a subject with a thyroid nodule is received. At 704, based on the received genetic variant data, a first polygenic risk score associated with thyroid cancer and a second polygenic risk score associated with benign thyroid nodular disease are determined. At 706, a composite risk metric is generated based on a combination of the first polygenic risk score and the second polygenic risk score, wherein the combination may reflect differential genetic contributions to malignant and benign thyroid nodular conditions. At 708, the thyroid nodule is classified asmalignant or benign based on the composite risk metric. At 710, an output indicative of the classification is generated, where the generated output may be used in clinical decision-making regarding the subject. In some examples, the output may comprise a recommendation regarding biopsy, surveillance, or surgical intervention. Flow 700 illustrates a structured approach for leveraging differential genetic contributions to malignant and benign thyroid nodular conditions to provide clinical guidance for thyroid nodule assessment.

[0056] FIG. 8 illustrates a flowchart depicting a flow 800 for classifying a thyroid nodule using combined genetic and ultrasound image-based assessment, in accordance with some aspects of the present disclosure. At 802, genetic variant data associated with a subject with a thyroid nodule is received. At 804, based on the received genetic variant data, a first polygenic risk score associated with thyroid cancer and a second polygenic risk score associated with benign thyroid nodular disease are determined. At 806, a composite risk metric is generated based on a weighted combination of the first polygenic risk score and the second polygenic risk score. At 808, the thyroid nodule is classified as malignant or benign based on the composite risk metric. At 810, an ultrasound image of the thyroid nodule is received. At 812, the ultrasound image is processed to generate an image-based risk assessment. At 814, the image-based risk assessment is combined with the composite risk metric to generate a combined classification of the thyroid nodule. At 816, an output indicative of the classification is generated, where the generated output may be used in clinical decision-making regarding the subject. The flow 800 illustrates a multimodal approach that leverages both genetic risk assessment through polygenic risk scores and image-based assessment through ultrasound analysis to provide a combined classification for thyroid nodule evaluation.

[0057] FIG. 9 illustrates a flowchart depicting a flow 900 for estimating a risk of aggressive thyroid cancer for a subject based on a polygenic risk score, in accordance with some examples of the present disclosure. At 902, the germline genetic data of a subject at risk for thyroid cancer is received. The flow 900 then proceeds to 904, a polygenic risk score for thyroid cancer is calculated based on a plurality of genetic variants associated with the subject. At 906, an association between the polygenic risk score and one or more aggressive thyroid cancer phenotypes is determined. In some examples, the aggressive thyroid cancer phenotypes may comprise one or more of tumor size, tumor focality, lymph node metastasis, extranodal extension, structural disease recurrence, or death from thyroid cancer. At 908, a risk ofaggressive thyroid cancer is estimated for the subject based on the polygenic risk score. At 910, an output indicative of the estimated risk of thyroid cancer is generated, where the generated output may be used in clinical decision-making regarding the subject. In some examples, the output may include a recommendation for targeted screening, earlier imaging, or preventative intervention. The flow 900 provides a structured approach for leveraging germline genetic information to identify individuals who may be at elevated risk for aggressive forms of thyroid cancer, enabling targeted screening or intervention strategies.

[0058] FIG. 10 illustrates a flowchart depicting a flow 1000 for estimating a risk of aggressive thyroid cancer for a subject based on a polygenic risk score, in accordance with some examples of the present disclosure. At 1002, germline genetic data of a subject at risk for thyroid cancer is received. At 1004, a polygenic risk score for thyroid cancer is calculated based on a plurality of genetic variants associated with the subject. At 1006, an association between the polygenic risk score and one or more aggressive thyroid cancer phenotypes is determined. At 1008, a risk of aggressive thyroid cancer for the subject is estimated based on the polygenic risk score. At 1010, the system determines whether the polygenic risk score exceeds a predefined threshold. In some examples, the predefined threshold may correspond to a top decile of a polygenic risk score distribution. If the polygenic risk score does not exceed the predefined threshold, the subject is not classified as a high risk for thyroid cancer, as in 1014. If the polygenic risk score exceeds the predefined threshold, the subject is classified as a high risk for thyroid cancer, as in 1012. At 1016, an output is generated that is indicative of the estimated risk of thyroid cancer, where the generated output may be used in clinical decision-making regarding the subject. In some examples, the output may comprise a recommendation for targeted screening, earlier imaging, or preventative intervention. The flow 1000 illustrates a structured approach for identifying individuals at elevated risk of aggressive thyroid cancer through genetic risk assessment, enabling targeted screening or preventative intervention based on the classification outcome.

[0059] FIG. 11 illustrates a flowchart depicting a flow 1100 for classifying a thyroid nodule using combined genetic and ultrasound image-based assessment, in accordance with some aspects of the present disclosure. At 1102, the germline genetic data of a subject at risk for thyroid cancer is received. At 1104, a polygenic risk score for thyroid cancer is calculated based on a plurality of genetic variants associated with the subject. At 1106, an association betweenthe polygenic risk score and one or more aggressive thyroid cancer phenotypes is determined. At 1108, a risk of aggressive thyroid cancer is estimated for the subject based on the polygenic risk score. At 1110, an ultrasound image of the thyroid nodule is received. At 1112, the ultrasound image is processed to generate an image-based risk assessment. At 1114, the imagebased risk assessment is combined with the polygenic risk score to generate a combined imagebased risk assessment of the thyroid nodule. At 1116, an output indicative of the estimated risk of thyroid cancer is generated, where the generated output may be used in clinical decisionmaking regarding the subject. In some examples, the output may comprise a recommendation regarding biopsy, surveillance, or surgical intervention. The flow 1100 illustrates the integration of genetic risk assessment with ultrasound image analysis to provide a comprehensive evaluation of thyroid nodule malignancy risk.

[0060] FIG. 12 illustrates a flowchart depicting a flow 1200 for post-GWAS quality control processing, in accordance with some aspects of the present disclosure. At 1202, GWAS summary data is received. At 1204, variants are filtered based on several criteria including: alleles including characters outside A, T, G, C; effect size (beta) values greater than le6 or less than -le6; standard error values greater than le6 or less than -le6; p-value equal to NA; imputation INFO score less than 0.3; allele frequency less than 0.0005 or greater than 0.9995; and allele count less than 40. At 1206, the filtered variants are harmonized to gnomAD, where the variants are aligned to the gnomAD reference database. At 1208, Allele frequency distance filtering is performed, which removes variants where the Mahalanobis distance between gnomAD allele frequency and harmonized GWAS allele frequency exceeds 3 standard deviations from the mean. At 1210, population structure analysis is performed. At 1212, continental ancestry proportions are determined with the population structure analysis being received as an input. At 1214, covariate-adjusted linkage disequilibrium score regression (Cov-LDSC) regression is performed using the results of population structure analysis and the determined continental ancestry proportions. The Cov-LDSC regression can be used to generate quality control metrics, as in 1216. The quality control metrics can include Cov-LDSC regression y-axis intercept, mean chi-squared, genomic inflation factor, and QQ and Manhattan plots. At 1218, analysis-ready GWAS data is produced. At 1220, meta-analysis on the data can be performed. In some examples, the flow 1200 provides a structured approach for processing and quality controlling GWAS summary data to ensure data integrity before meta-analysis.

[0061] Tn some examples, genomic loci may be defined using a linkage disequilibrium (LD) clumping procedure. In an illustrative example, the LD clumping procedure may use an index variant significance threshold of P < 5 x 10 \ a search window of approximately 5 Mb, and an r2threshold of approximately 0.01. Independent clumps may be defined as those that do not share variants associated with the phenotype at P < 1x10?. One variant with the lowest P value from each independent clump may be selected as a lead variant for downstream analysis. These parameters may help ensure that the identified loci represent independent genetic signals contributing to thyroid disease risk.

[0062] In some examples, a locus may be classified as novel or previously discovered based on comparison to existing genetic association databases. In an illustrative example, a locus may be considered novel if no variants for the corresponding phenotype have been reported within ±500 kb in a reference database such as the NHGRI-EBI GWAS Catalog. Otherwise, the variant may be labeled as previously discovered. This classification approach may help distinguish newly identified genetic associations from those that replicate prior findings, providing context for the contribution of the meta-analysis to the understanding of thyroid disease genetics.

[0063] In some examples, variant filtering may be performed to minimize false positive associations introduced by confounding within a single large biobank. In an illustrative example, only variants present in at least four input GWAS datasets may be considered in the downstream analysis. If three or fewer datasets are available for an ancestry-stratified meta-analysis, the threshold may be set to two datasets. Heterogeneity across datasets may be assessed by calculating Cochran's Q P values. This multi-biobank consistency requirement may help ensure that the identified genetic associations are robust and replicable across independent study populations.

[0064] In some examples, the thyroid cancer cases in the validation cohort may comprise multiple histological subtypes. For example, papillary thyroid cancer may be the most common thyroid malignancy, followed by follicular thyroid carcinoma, oncocytic thyroid carcinoma, anaplastic thyroid carcinoma and poorly differentiated thyroid carcinoma. The PRS may be applicable across these thyroid cancer subtypes, supporting broad clinical utility.

[0065] In some examples, the PRS may provide a noninvasive risk assessment that is independent of somatic changes used in molecular tests for the management of thyroid nodules with indeterminate cytology. Therefore, incorporating PRS into existing molecular diagnostictests may improve their performance. In some examples, the PRS may guide which biopsy-proven thyroid cancers are likely to be indolent and therefore suitable for active surveillance. In some examples, the PRS may improve the assessment of the postoperative risk of recurrence and inform postoperative management decisions.

[0066] In some implementations, the PRS, composite risk metric, image-based risk assessment, or combined PRS+image-based artificial intelligence(AI) classification is integrated with a molecular diagnostic test for an indeterminate thyroid nodule. For example, the system may combine inherited genetic risk with somatic mutation status, gene expression classifier output, copy-number alteration data, methylation data, fusion detection, microRNA expression, or other molecular diagnostic features. The combined output may reclassify an indeterminate thyroid nodule as likely benign, likely malignant, low risk, intermediate risk, high risk, suitable for observation, suitable for repeat FNA, suitable for molecular testing, or suitable for surgery.

[0067] In some examples, the genetic risk assessment may be further improved by incorporating various types of genetic alteration beyond common variants, such as rare variants and syndromic mutations associated with hereditary thyroid cancer syndromes. These additional genetic alterations may further improve the understanding of the biology of thyroid disease and improve risk assessments for clinical applications.

[0068] Although whole-genome sequencing is described in certain examples, the genetic variant data may be obtained from any suitable germline genetic assay, including whole-genome sequencing, whole-exome sequencing, targeted sequencing, SNP array genotyping, imputed genotype data, low-pass sequencing, genotyping-by-sequencing, clinical genetic testing, or combinations thereof. The genetic variant data may comprise genotype calls, allele dosages, imputed dosages, variant probabilities, variant presence / absence calls, pathogenic or likely pathogenic variant calls, rare variant calls, common variant calls, or ancestry-informative markers.

[0069] In some implementations, the genetic risk assessment combines a common-variant polygenic risk score with one or more rare variants, pathogenic variants, likely pathogenic variants, syndromic thyroid cancer variants, or variants in actionable cancer-predisposition genes. The rare-variant or pathogenic-variant information may be encoded as a binary feature, weighted score, burden score, gene-level feature, pathway-level feature, or covariate andcombined with the PRS, composite risk metric, image-based risk assessment, or meta-classifier output.

[0070] FIG. 13 illustrates a block diagram of a system 1300 for classifying a thyroid nodule, in accordance with some aspects of the present disclosure. The system 1300 includes a processor 1302 and a memory 1304. The memory 1304 stores genetic variant data 1306, germline genetic data 1308, and ultrasound image data 1310. The system 1300 also includes a first polygenic risk score module 1312, a second polygenic risk score module 1314, a composite risk metric module 1316, a classification module 1318, an ultrasound image processing module 1320, and a polygenic risk score module 1322.

[0071] The polygenic risk score module 1312 can be configured to determine a first polygenic risk score associated with thyroid cancer and the second polygenic risk score module 1314 configured to determine a second polygenic risk score associated with benign thyroid nodular disease. The composite risk metric module 1316 can be configured to generate a composite risk metric based on a combination of the first polygenic risk score and the second polygenic risk score. The classification module 1318 can be configured to classify the thyroid nodule as malignant or benign based on the composite risk metric. The ultrasound image processing module 1320 can be configured to process the ultrasound image data 1310 using a trained deep learning model to generate an image-based risk assessment. The polygenic risk score module 1322 can be configured to determine a polygenic risk score for thyroid cancer based on a plurality of genetic variants associated with a subject. The system 1300 further includes a display 1324 for presenting output indicative of the classification for use in clinical decision-making. In some examples, the display 1324 may present recommendations regarding biopsy, surveillance, or surgical intervention as described in the clinical trial protocol. The system 1300 also includes a communication interface 1326 for data exchange with external devices or networks, an input / output device 1328 for receiving user input and providing system output, and a machine readable medium 1330 storing instructions executable by the processor 1302. The processor 1302 executes instructions stored in the memory 1304 to coordinate the operation of the various modules and generate recommendations for clinical decision-making regarding thyroid nodule assessment.

[0072] FIG. 14 illustrates a block diagram of a computer model system 1400 depicting computer model inference and training operations, in accordance with some aspects of thepresent disclosure. The computer model system 1400 includes two operational sections: a computer model inference section and a computer model training section. In the computer model inference section, input data 1404 is provided to a computer model 1402. The computer model 1402 includes model parameters 1408 that define the behavior and characteristics of the model. The computer model 1402 processes the input data 1404 using the model parameters 1408 to generate a model output 1406. In some aspects, the computer model 1402 in the inference section corresponds to the trained deep learning model used by the ultrasound image processing module 1320 of system 1300 to generate image-based risk assessments. In some examples, the input data 1404 may comprise ultrasound image data of thyroid nodules, such as the static images and video clips described in Section 1.3. In some examples, the model output 1406 may include an imagebased risk assessment indicating a probability of malignancy or a classification of the thyroid nodule as benign, papillary thyroid cancer-like, or follicular thyroid cancer-like. In the computer model training section, training data 1412 provides training inputs to a computer model 1410. The computer model 1410 also includes model parameters 1408 that are subject to adjustment during the training process. The computer model 1410 processes the training inputs to generate model predicted outputs. Training data labels are provided to a model evaluation 1414 component along with the model predicted outputs. The model evaluation 1414 includes training parameters 1416 that govern the training process. The model evaluation 1414 compares the model predicted outputs against the training data labels and generates a parameter update signal that is fed back to the computer model 1410 to adjust the model parameters 1408. In some examples, the training data 1412 may include the 152,300 images and 1534 cine clips from the 1516 thyroid nodules described in Section 1.3. In some examples, the training data labels may include benign or malignant histopathology results used to train the deep learning model to recognize thyroid cancer. The computer model system 1400 illustrates the relationship between inference operations, where a trained model generates outputs from new input data, and training operations, where the model parameters 1408 are optimized based on labeled training data through an iterative evaluation and update process.

[0073] As described above in connection with FIG. 14, the computer model 1402 generates the model output 1406 based on the model parameters 1408. The output of the computer model 1402 may be referred to as an 'inference' because it is a predictive value based on the input data 1404 and based on previous example data used in the model training. The parameters of themodel may be determined based on a training process that finds an optimization of the model parameters 1408, typically using training data and desired outputs of the model for the respective training data as discussed below.

[0074] In addition to the ultrasound image data and classification outputs described in connection with FIG. 14, the input data 1404 and the model output 1406 may vary according to the particular use case. In other applications, the input data 1404 may include pixel data extracted from DICOM format ultrasound studies. In addition, the input data 1404 may be a processed version of another type of input object, for example, representing various features of the ultrasound image or representing preprocessing of the ultrasound image before input to the computer model 1402. As one example, an ultrasound image may be processed to extract a bounding box region corresponding to the thyroid nodule, which is the input data 1404 processed by the computer model 1402. In some examples, the input data 1404 may also include genetic variant data or polygenic risk score values associated with a subject, which may be combined with image-based features in a meta-classifier to produce a combined PRS+AI classification. Although not further discussed here, such further computer models may be independently or jointly trained with the computer model 1402.

[0075] The computer model 1402 includes various model parameters 1408, as noted above, that describe the characteristics and functions that generate the model output 1406 from the input data 1404. In particular, the model parameters 1408 may include a model structure, model weights, and a model execution environment. The model structure may include, for example, the particular type of computer model 1402 and its structure and organization. For example, the model structure may designate a neural network, which may be comprised of multiple layers, and the model parameters 1408 may describe individual types of layers included in the neural network and the connections between layers (e.g., the output of which layers constitute inputs to which other layers). Such networks may include, for example, feature extraction layers, convolutional layers, pooling / dimensional reduction layers, activation layers, output / predictive layers, and so forth. While in some instances the model structure may be determined by a designer of the computer model, in other examples, the model structure itself may be learned via a training process and may thus form certain “model parameters” of the model. In some examples, the model structure may designate a three-dimensional convolutional neural networkfor processing ultrasound video clips or a vision transformer architecture such as SLIViT for extracting features from three-dimensional biomedical imaging data.

[0076] The model weights may represent the values with which the computer model 1402 processes the input data 1404 to the model output 1406. Each portion or layer of the computer model 1402 may have such weights. For example, weights may be used to determine values for processing inputs to determine outputs at a particular portion of a model. Stated another way, for example, model weights may describe how to combine or manipulate values of the input data 1404 or thresholds for determining activations as output for a model. As one example, a convolutional layer typically includes a set of convolutional “weights,” also termed a convolutional kernel, to be applied to a set of inputs to that layer. These are subsequently combined, typically along with a “bias” parameter, and weights for other transformations to generate an output for the convolutional layer.

[0077] The model execution parameters represent parameters describing the execution conditions for the model. In particular, aspects of the model may be implemented on various types of hardware or circuitry for executing the computer model 1402. For example, portions of the model may be implemented in various types of circuitry, such as general-purpose circuity (e.g., a general CPU), circuity specialized for certain functions (e.g., a GPU or programmable Multiply-and-Accumulate circuit) or circuitry specially designed for the particular computer model application. In some configurations, different portions of the computer model 1402 may be implemented on different types of circuitries. As discussed below, training of the model may include optimizing the types of hardware used for certain aspects of the computer model 1402 (e.g., co-trained), or may be determined after other parameters for the computer model 1402 are determined without regard to configuration executing the model. In another example, the execution parameters may also determine or limit the types of processes or functions available at different portions of the model, such as value ranges available at certain points in the processes, operations available for performing a task, and so forth.

[0078] Computer model training may thus be used to determine or “train” the values of the model parameters 1408 for the computer model 1410. During training, the model parameters 1408 are optimized to “learn” values of the model parameters (such as individual weights, activation values, model execution environment, etc.), that improve the model parameters 1408 based on an optimization function that seeks to improve a cost function (also sometimes termeda loss function). Before training, the computer model 1410 has model parameters 1408 that have initial values that may be selected in various ways, such as by a randomized initialization, initial values selected based on other or similar computer models, or by other means. During training, the model parameters are modified based on the optimization function to improve the cost / loss function relative to the prior model parameters.

[0079] In many applications, training data 1412 includes a data set to be used for training the computer model 1410. The data set varies according to the particular application and purpose of the computer model 1410. In supervised learning tasks, the training data 1412 typically includes a set of training data labels that describe the training data 1412 and the desired output of the model relative to the training data 1412. In some examples, the training data labels may classify nodules into three classes comprising benign, papillary thyroid cancer-like, and follicular thyroid cancer-like, enabling the computer model 1410 to learn to distinguish between these classes.

[0080] In some implementations, training labels for the image-based model or meta-classifier are derived from Bethesda cytology category, molecular diagnostic result, surgical histopathology, benign follow-up, thyroid cancer subtype, papillary thyroid cancer-like status, follicular thyroid cancer-like status, oncocytic thyroid carcinoma status, follicular adenoma status, structural disease recurrence, lymph-node metastasis, extrathyroidal extension, extranodal extension, tumor size, tumor focality, or death from thyroid cancer. Model performance may be evaluated using sensitivity, specificity, AUC, positive predictive value, negative predictive value, calibration, decision-curve analysis, confusion matrix, or reduction in false-positive FNA recommendations.

[0081] To train the computer model 1410, a training module (not shown) applies the training inputs to the computer model 1410 to determine the outputs predicted by the model for the given training inputs. The training module, though not shown, is a computing module used for performing the training of the computer model 1410 by executing the computer model 1410 according to its inputs and outputs given the model’s parameters and modifying the model parameters based on the results. The training module may apply the actual execution environment of the computer model 1410, or may simulate the results of the execution environment, for example to estimate the performance, runtime, memory, or circuit area (e.g., if specialized hardware is used) of the computer model 1410. The training module, along with thetraining data 1412 and model evaluation, may be instantiated in software and / or hardware by one or more processing devices. In various examples, the training process may also be performed by multiple computing systems in conjunction with one another, such as distributed / cloud computing systems.

[0082] After processing the training inputs according to the current model parameters for the computer model 1410, the model’s predicted outputs are evaluated and the computer model 1410 is evaluated with respect to the cost function and optimized using an optimization function of the training model. Depending on the optimization function, particular training process and training parameters 1416 after the model evaluation are updated to improve the optimization function of the computer model 1410. In supervised training (i.e., training data labels are available), the cost function may evaluate the model’s predicted outputs relative to the training data labels and to evaluate the relative cost or loss of the prediction relative to the “known” labels for the data. This provides a measure of the frequency of correct predictions by the computer model 1410 and may be measured in various ways, such as the precision (frequency of false positives) and recall (frequency of false negatives). In some examples, for the thyroid nodule classification task, the cost function may evaluate the sensitivity and specificity of the model's predictions relative to histopathology-confirmed diagnoses. The cost function in some circumstances may also evaluate other characteristics of the model, for example the model complexity, processing speed, memory requirements, physical circuit characteristics (e.g., power requirements, circuit throughput) and other characteristics of the computer model 1410 structure and execution environment (e.g., to evaluate or modify these model parameters).

[0083] After determining results of the cost function, the optimization function determines a modification of the model parameters to improve the cost function for the training data 1412. Many such optimization functions are known to one skilled on the art. Many such approaches differentiate the cost function with respect to the parameters of the model and determine modifications to the model parameters that thus improves the cost function. The parameters for the optimization function, including algorithms for modifying the model parameters are the training parameters 1416 for the optimization function. For example, the optimization algorithm may use gradient descent (or its variants), momentum-based optimization, or other optimization approaches used in the art and as appropriate for the particular use of the model. The optimization algorithm thus determines the parameter updates to the model parameters. In someimplementations, the training data 1412 is batched and the parameter updates are iteratively applied to batches of the training data 1412. For example, the model parameters may be initialized, then applied to a first batch of data to determine a first modification to the model parameters. The second batch of data may then be evaluated with the modified model parameters to determine a second modification to the model parameters, and so forth, until a stopping point, typically based on either the amount of training data 1412 available or the incremental improvements in model parameters are below a threshold (e.g., additional training data 1412 no longer continues to improve the model parameters). Additional training parameters 1416 may describe the batch size for the training data 1412, a portion of training data 1412 to use as validation data, the step size of parameter updates, a learning rate of the model, and so forth. Additional techniques may also be used to determine global optimums or address nondifferentiable model parameter spaces.

[0084] FIG. 15 illustrates a block diagram of a neural network architecture for thyroid nodule classification, in accordance with some examples of the present disclosure. The neural network comprises an input layer 1502, one or more hidden layer(s) 1504, and an output layer 1506 arranged in a sequential configuration. In some examples, the neural network architecture may be used to implement the trained deep learning model of the ultrasound image processing module 1320 of system 1300 described in connection with FIG. 13, the computer model 1402 described in connection with FIG. 14, and the trained deep learning model 1604 described in connection with FIG. 16. In some examples, the input layer 1502 may receive ultrasound image data of thyroid nodules, such as static images or video clips. The hidden layer(s) 1504 may extract features from the ultrasound image data through successive transformations, and may implement convolutional layers, pooling layers, or attention mechanisms suitable for the particular input data type. The output layer 1506 may generate an image-based risk assessment, such as a probability of malignancy or a classification of the thyroid nodule into one of at least three classes comprising benign 1606, papillary thyroid cancer (PTC)-like 1608, and follicular thyroid cancer (FTC)-like 1610 as described in connection with FIG. 16.

[0085] The values for data in each layer of the network are generally determined based on one or more prior layers of the network. Each layer of a network generates a set of values, termed “activations” that represent the output values of that layer of a network and may be the input to the next layer of the network. For the input layer 1502, the activations are typically the valuesof the input data, although the input layer 1502 may represent input data as modified through one or more transformations to generate representations of the input data. For example, ultrasound image data of a thyroid nodule may be represented as pixel values in a matrix corresponding to the image resolution. The input layer 1502 may receive this image data directly or as a transformation of the raw pixel data. The neural network may also receive the output of another computer model (or several), as its input layer 1502, such that the input layer 1502 of the neural network shown in FIGURE 15 is the output of another computer model. Accordingly, each layer may receive a set of inputs, also termed “input activations,” representing activations of one or more prior layers of the network and generate a set of outputs, also termed “output activations” representing the activation of that layer of the network. Stated another way, one layer’s output activations become the input activations of another layer of the network, except for the final output layer of 1506 of the network.

[0086] Each layer of the neural network typically represents its output activations (i.e., also termed its outputs) in a matrix, which may be 1, 2, 3, or n-dimensional according to the particular structure of the network. As shown in FIGURE 15, the dimensionality of each layer may differ according to the design of each layer. The dimensionality of the output layer 1506 depends on the characteristics of the prediction made by the model. For example, a computer model for multi-object classification may generate an output layer 1506 having a one-dimensional array in which each position in the array represents the likelihood of a different classification for the input layer 1502. In another example for classification of portions of an image, the input layer 1502 may be an image having a resolution, such as 512x512, and the output layer may be a 512x512xn matrix in which the output layer 1506 provides n classification predictions for each of the input pixels, such that the corresponding position of each pixel in the input layer 1502 in the output layer 1506 is an n-dimensional array corresponding to the classification predictions for that pixel.

[0087] The hidden layers 1504 provide output activations that variously characterize the input layer 1502 in various ways that assist in effectively generating the output layer 1506. The hidden layers thus may be considered to provide additional features or characteristics of the input layer 1502. Though two hidden layers are shown in FIGURE 15, in practice any number of hidden layers may be provided in various neural network structures.

[0088] Each layer generally determines the output activation values of positions in its activation matrix based on the output activations of one or more previous layers of the neural network (which may be considered input activations to the layer being evaluated). Each layer applies a function to the input activations to generate its activations. Such layers may include fully-connected layers (e.g., every input is connected to every output of a layer), convolutional layers, deconvolutional layers, pooling layers, and recurrent layers. Various types of functions may be applied by a layer, including linear combinations, convolutional kernels, activation functions, pooling, and so forth. The parameters of a layer’s function are used to determine output activations for a layer from the layer’s activation inputs and are typically modified during the model training process. The parameters describing the contribution of a particular portion of a prior layer is typically termed a weight. For example, in some layers, the function is a multiplication of each input with a respective weight to determine the activations for that layer. For a neural network, the parameters for the model as a whole thus may include the parameters for each of the individual layers and in large-scale networks can include hundreds of thousands, millions, or more of different parameters.

[0089] As one example for training a neural network, the cost function is evaluated at the output layer 1506. To determine modifications of the parameters for each layer, the parameters of each prior layer may be evaluated to determine respective modifications. In one example, the cost function (or “error”) is backpropagated such that the parameters are evaluated by the optimization algorithm for each layer in sequence, until the input layer 1502 is reached.

[0090] FIG. 16 illustrates a block diagram of a deep learning model classification architecture for thyroid nodule assessment, in accordance with some aspects of the present disclosure. Ultrasound image data 1602 is provided as input to a trained deep learning model 1604. The trained deep learning model 1604 processes the ultrasound image data 1602 and generates a classification output indicating one of three classes: benign 1606, PTC-like 1608, or FTC-like 1610. The benign 1606 classification corresponds to thyroid nodules determined to be non-malignant, which may include hyperplastic nodules and adenomas. The PTC-like 1608 classification corresponds to papillary thyroid cancer-like nodules, which may include classic papillary thyroid cancers and tall cell variants that typically exhibit suspicious sonographic features such as hypoechogenicity and microcalcifications. The FTC-like 1610 classification corresponds to follicular thyroid cancer-like nodules, which may include follicular and oncocyticthyroid carcinomas that frequently lack the suspicious sonographic features associated with papillary thyroid cancers. In some examples, the trained deep learning model 1604 may correspond to the ultrasound image processing module 1320 of system 1300 described in connection with FIG. 13 and the neural network architecture described in connection with FIG.15. The three-class classification approach addresses the deficiency of machine learning tools trained to recognize only two classes, benign and malignant, and improves the ability to identify difficult-to-diagnose thyroid cancers such as follicular and oncocytic thyroid carcinomas as described in Section 1.3. In some examples, the output from the trained deep learning model 1604 may be combined with a polygenic risk score via a meta-classifier to produce a combined PRS+AI classification for use in clinical decision-making regarding biopsy, surveillance, or surgical intervention.

[0091] Having described the figures and their associated elements, the following sections provide additional detail regarding embodiments and examples of the inventive concept, including the genetic architecture underlying the polygenic risk scores, the development and validation of the combined genetic and image-based classifier, and the clinical trial protocol for evaluating the classifier in a prospective clinical setting.

[0092] Embodiments of the present inventive concept are intended to work in patients with thyroid nodules that meet the criteria for fine needle aspiration (FNA) as per the current standard of care. Embodiments of the present inventive concept identify thyroid nodules that are very unlikely to be malignant and can be managed with observation. The system prevents unnecessary procedures, such as thyroid nodule biopsies and diagnostic thyroidectomies, in patients with benign thyroid nodules.

[0093] In some embodiments, the inventive concept works as follows:1. The patient undergoes a blood draw, genomic DNA extraction and whole genome sequencing.2. Hereditary predisposition to developing thyroid cancer is estimated based on the presence of 88 or more genetic variants known to increase the risk of thyroid cancer (polygenic risk score). This list of variants was derived from the very large global metaanalysis of genome-wide association studies for all common thyroid diseases performed by our team. The list of variants may be updated as new genetic associations with thyroid cancer and other thyroid diseases are discovered.3. In parallel with genetic risk assessment, thyroid nodule ultrasound images are analyzed by the neural network trained to recognize thyroid cancer.4. The neural network alone performs at the level of a senior radiologist. However, when combined with genetic risk assessment, the system achieves an outstanding performance with a sensitivity of 95% and a specificity of 63% on retrospective analysis.5. More than half of the patients who would require FNA per the current standard of care are now recommended for observation.EXAMPLES

[0094] An example of the inventive concept includes:1. Genetic variants associated with diseases are discovered with the help of genomewide association studies (GWAS). To better define the genetic architecture of thyroid cancer and related thyroid diseases, the Virtual Thyroid Biopsy Consortium under the Global Biobank Meta-analysis initiative (GBMI, https: / / www.globalbiobankmeta.org) umbrella was assembled. The Consortium more than tripled the discovery power of the previous GWAS, aggregating data from 21,816 individuals with thyroid cancer and 2,895,812 controls from nineteen biobanks, ten countries and four continents. The unmatched statistical power led to the discovery of hundreds of novel genetic associations with thyroid cancer, benign nodular goiter, Graves' disease and hypothyroidism.2. The newly defined genetic architecture for thyroid cancer and benign nodular goiter was used to develop a polygenic risk score specifically designed to distinguish benign and malignant thyroid nodules.3. The polygenic risk score derived from the global GWAS meta-analysis described herein was found to significantly outperform scores derived from the largest public GWAS meta-analysis for thyroid cancer and to improve discrimination of malignant versus benign thyroid nodules. For example, PRS ThC vs. All achieved an AUC of about 0.622, and PRS ThC vs. BNG improved performance for the clinical thyroid-cancer-versus-benign- nodular-goiter task, with an AUC of about 0.670.4. A deep learning ultrasound image-based thyroid nodule classifier was trained using 32,545 images from 621 nodules. The model is retrained using a much larger training image set that is collected.5. The combined polygenic risk score and image-based classifier was tested in a proof-of-concept study and demonstrated synergistic performance of genetic and imagebased assessments.

[0095] Ways to improve the combined genetic and image-based classifier performance are tested, including further increasing GWAS meta-analysis statistical power, leveraging known pathogenic and likely pathogenic variants for syndromic thyroid cancer, and using, for example, thyroid nodule ultrasound video clips for improved prediction.

[0096] Evaluating thyroid nodules to rule out malignancy is a common clinical task. Thyroid ultrasound data reveal nodules in up to 65% of the general population, with prevalence increasing in women, with age, and after radiation exposure (1,2). Clinical providers assess thyroid nodule sonographic characteristics and use ultrasound-based risk stratification algorithms like the American College of Radiology Thyroid Imaging Reporting and Data System (TI-RADS) (3) to decide if fine needle aspiration (FNA) biopsy is indicated to rule out cancer. Over 600,000 FNAs are performed annually in the United States (4) at an estimated cost of $480 - $720 million.

[0097] Despite the wide adoption of TI-RADS, only 8% of biopsy results meet cytologic categories requiring thyroid surgery; 72% are benign, and the remaining 20% are inadequate or indeterminate, necessitating repeat procedures, expensive molecular testing, or diagnostic surgery (5-8). More than half of diagnostic thyroidectomies performed in patients with indeterminate thyroid nodules reveal benign pathology (9). Molecular diagnostic tests for somatic oncogenic mutations and high-risk gene expression patterns (10,11) reduce the need for diagnostic thyroidectomies, but the positive predictive value of these tests remains low (47-68%) (12,13). Improving early assessment of thyroid cancer risk in patients with thyroid nodules is critical to reduce unnecessary procedures, overtreatment, excess costs to the health system and patient harm.

[0098] Population-based registries demonstrate that thyroid cancer is highly heritable, but the genetic components are not well-defined. Genetic effects contribute 28-53% to thyroid cancer susceptibility (14,15). The thyroid cancer germline genetic risk assessment according to implementations and embodiments of the inventive concept is complementary and synergistic to the current standard of care image-based risk stratification.

[0099] Hereditary risk is quantified using an ancestry-aware polygenic risk score (PRS) to help improve the identification of thyroid nodules requiring FNA and to support riskstratification for aggressive thyroid cancer. This is based on a global multi-ancestry metaanalysis of genome-wide association studies (GWAS) aggregating genotype data from 2.9 million individuals from 19 biobanks in 10 countries and four continents. The meta-analysis identified mechanistically plausible genetic associations in genes with known roles in thyroid cancer development and aggressiveness. The polygenic risk score (PRS) developed from meta-analysis demonstrated excellent performance in identifying patients at risk for thyroid cancer (area under the receiver operating characteristic curve [AUC] of 0.686 [95% CI, 0.669, 0.703]). Notably, the PRS prediction accuracy is comparable in subjects of diverse genetic ancestries, including EUR-like, AFR-like, AMR-like, and EAS-like populations, ensuring broad applicability.

[0100] In some examples, the meta-analysis can aggregate genetic data from individuals of diverse genetic ancestries. The population structure may be determined via mixture modeling of study -based allele frequencies compared to a reference panel. In an illustrative example, approximately 17% of genotypes can be from individuals of African (AFR-like) ancestry, approximately 4.4% from Admixed American (AMR-like) ancestry, approximately 8.1% from East Asian (EAS-like) ancestry, and approximately 70.5% from European (EUR-like) ancestry. This multi-ancestry composition supports the development of polygenic risk scores that perform equitably across diverse populations.

[0101] Risk factors other than genetics are known to impact thyroid cancer risk, suggesting that PRS should be used alongside other clinical approaches. PRS is combined with state-of-the-art thyroid nodule ultrasound-based imaging assessment to model the probability of malignancy. A first-generation deep learning thyroid nodule classifier (16) trained on ultrasound images performs comparably to a senior radiologist. It achieved an AUC of 0.83, which improved to 0.92 when combined with PRS, genetic ancestry and demographic covariates. In a proof-of-concept study (16), it was demonstrated that combined genetic and image-based risk assessment outperforms current clinical standard-of-care risk stratification schemas, reducing the need for FNA while maintaining good sensitivity for detecting thyroid cancer.

[0102] Foundational work of this translational study indicates that the clinical implementation of a multimodal image-based and genetic thyroid nodule classifier can greatly reduce the number of unnecessary thyroid nodule FNA biopsies. It is estimated that using such a system can reduce the need for FNA by up to 175,000 annually in the United States. Moreimportantly, fewer biopsies translate into fewer indeterminate cytology results and reduce the need for molecular testing, diagnostic thyroidectomies, patient impact and healthcare costs.CLASSIFIER DEVELOPMENT AND ASSESSMENT

[0103] An accurate, equitable and interpretable clinical system for thyroid nodule risk assessment is developed. The clinical implementation of a multimodal image-based and genetic thyroid nodule classifier reduces the number of unnecessary thyroid nodule fine needle aspiration biopsies (FNA) as indicated in the following disclosure.1. A combined genetic and image-based thyroid nodule classifier for thyroid cancer diagnosis is provided.

[0104] Prior work is built upon to develop a clinical polygenic risk score (PRS) and ultrasound image-based deep learning model that estimates the probability of cancer in thyroid nodules. This provides: 1) maximized performance (increase specificity and positive predictive value while maintaining the sensitivity for detecting thyroid cancer comparable to that of FNA [>90%]); 2) ensured equitable performance in diverse populations; 3) ensured performance across thyroid cancer subtypes, and 4) provided risk assessment understandable to clinical providers.1.1. Genetic architecture of thyroid diseases.

[0105] Thyroid cancer heritability is high, with up to 53% of risk determined by germline genetic factors (14,17,18). Germline genetic risk assessment presents a significant opportunity to improve the diagnosis of thyroid cancer, and it is not yet used clinically (with exceptions, e.g., testing for syndromic thyroid cancers).

[0106] Genome-wide association studies (GWAS) help discover genetic variants that predispose individuals to the diseases. The Virtual Thyroid Biopsy Consortium was assembled under the Global Biobank Meta-analysis initiative (GBMI, https: / / www.globalbiobankmeta.org) umbrella to better define the genetic architecture of thyroid cancer and related thyroid diseases. The Consortium more than tripled the discovery power of previous GWAS, aggregating data from 21,816 individuals with thyroid cancer and 2,895,812 controls from 19 biobanks, 10 countries and four continents. The unmatched statistical power enabled the discovery of hundreds of novel genetic associations with thyroid cancer, benign nodular goiter, Graves' disease and hypothyroidism. Importantly, many variants associated with thyroid cancer werewithin genes and pathways known to be associated with malignant transformation and thyroid cancer: cell cycle, telomere maintenance, apoptosis and ETS family transcription factors. Four genes with significant germline variants increasing thyroid cancer risk (TERT, TP53, DLG5 and CHEK2) are also somatically altered in thyroid cancers (19,20).

[0107] The meta-analysis identified 883 independent loci significantly associated with thyroid diseases, including 313 variants previously reported to the NHGRI-EBI Catalog for thyroid traits and 570 new loci. Among 36 loci associated with thyroid cancer (ThC) but not benign nodular goiter (BNG), seven are in genes that encode components of cell cycle checkpoints, proteins regulating centrosome and kinetochore function, microtubule attachment and chromosome segregation (CDCA7L, CENPE, CEP 120, CHEK2, NUF2, PMF1, TP53). The ThC-specific locus Cllorf65 overlaps with the cell cycle checkpoint kinase gene ATM (for example, 11:108267276, ATM p.Phe858Leu), which is frequently mutated in advanced ThC. Five genes with ThC-specific associations have a role in DNA repair and cellular response to DNA damage (A TM, DCLRE1B, PCNX2, EX01, TP 53).

[0108] In some examples, the lead variants identified in the meta-analysis can be classified by genomic location. In an illustrative example, most lead variants were intronic (for example, approximately 407 variants), followed by intergenic variants (for example, approximately 302 variants). Among significant exonic variants, a majority may be nonsynonymous, potentially altering protein function. For example, among 46 significant exonic variants, 43 were nonsynonymous. This distribution of variant types can provide additional context for the role of the identified genetic loci in thyroid disease susceptibility.

[0109] BNG-specific loci are located in genes participating in insulin-like growth factor 1 (IGF1 and IGF2BP2) and fibroblast growth factor (FGF7 an FRS2) signaling pathways. Genes having a role in thyroid gland development and thyroid hormone synthesis were linked to benign nodules (GLIS3, TPO), but some are also associated with ThC (NKX2-1, TG). Notably, telomere maintenance genes (ACTRT3, LRRC6, STN1, TERT) were associated with both ThC and BNG. Of particular interest is the PTCSC2 locus because its significant variants have the opposite direction of effect with ThC and BNG (p = -0.77, P = 1.2 * 1024), representing the only locus identified in the GWAS meta-analysis where significant variants display inversely associated effects between malignant and benign thyroid nodular disease.

[0110] Tn some implementations, a paradigm explaining genetic predisposition to benign and malignant thyroid nodules is proposed. Two biological processes with distinct genetic architecture may cause thyroid nodules: (1) hyperplasia, a polyclonal follicular cell proliferation with no malignant potential; and (2) neoplasia, a clonal growth driven by somatic genetic alterations. Neoplastic nodules can be benign or malignant, and the mismatch between biological mechanisms (hyperplasia versus neoplasia) and GWAS phenotype definitions (benign and malignant thyroid nodules) has led to apparent genetic pleiotropy. Genes participating in the cell cycle, DNA repair and cellular response to DNA damage are predominantly associated with ThC but not benign nodules, highlighting the importance of these biological processes for malignant transformation of thyroid follicular cells. Genes in the fibroblast growth factor and insulin-like growth factor 1 signaling pathways are uniquely associated with BNG and may lead to hyperplastic benign thyroid nodules without malignant potential. Variants in genes participating in telomere maintenance increase the risk of ThC and benign neoplastic thyroid nodules (adenomas).[OHl] Gene expression analysis evaluated the mRNA expression of genes discovered in the ThC GWAS meta-analysis and the TWAS in normal and malignant thyroid tissues. Of 20 evaluated tissues, normal thyroid tissue was among the top three highest-expressing tissues for 20 genes. Two genes, TG and NKX2-1, are expressed only in the thyroid. The expression of six genes (ETS1, HMGA2, NFIA, PCNX2, PIBF1 and VAV ) significantly correlated with younger age at ThC diagnosis. TERT expression correlated with older age at diagnosis, matching a similar association with somatic TERT promoter mutations. The expression of 23 genes was positively correlated with at least one clinical or molecular ThC risk feature: younger age at diagnosis, higher stage, presence of extrathyroidal extension, lower BRAF / RAS score (indicating a BRAF-like expression profile), higher ERK score (measuring RAS / MAPK pathway activity) and lower ThC differentiation.

[0112] Ancestry-stratified GWAS meta-analyses in European (EUR), African American (AFR), East Asian (EAS) and Admixed American (AMR) populations were performed, uncovering additional associations with thyroid cancer and benign thyroid diseases and improving trans-ancestry performance of polygenic risk scores (Section 1.2). This is the only adequately powered study that systematically investigates thyroid cancer genetics in non-European individuals. This diversity improves PRS portability and ensures personalized insights for broader populations (21).

[0113] Significant genetic correlations were found between thyroid diseases using covariate-adjusted linkage disequilibrium score regression. There was a strong genetic correlation between lymphocytic thyroiditis (LT) and hypothyroidism (rg= 0.97). Significant genetic correlations were also found between LT and Graves disease (GD) (rg= 0.62), ThC and BNG (rg= 0.41), GD and hypothyroidism (rg= 0.37), GD and BNG (rg= 0.31), and GD and ThC (rg= 0.20). These genetic correlations identify physiologically plausible and clinically meaningful associations between thyroid diseases. The shared genetic basis for LT and GD is expected because both conditions are autoimmune diseases with highly concordant familial risk. The genetic correlation between GD and thyroid nodular disease (both benign nodules and ThC) is mechanistically explained by enhanced TSH receptor signaling, which promotes thyroid epithelial growth and protects thyroid cells from apoptosis.

[0114] Overall, global collaboration and inclusivity significantly increased the understanding of the genetic architecture of thyroid diseases, enabling the development of a clinically relevant PRS for thyroid cancer diagnosis.

[0115] In some examples, certain loci may be discovered in Graves' disease (GD) metaanalysis but not in the hypothyroidism GWAS meta-analysis, despite the greater discovery power of the hypothyroidism GWAS. In an illustrative example, nine loci, including CD40, LINGO2, TNRC18, and TERT, are significantly associated with GD but not with hypothyroidism. These genes are involved in immune system processes that define the type of autoantibodies produced: TSH receptor antibodies in GD or thyroid peroxidase / thyroglobulin antibodies in lymphocytic thyroiditis and primary hypothyroidism. Consistently, variants in TSHR are strongly associated with GD, while TRO and TG associations may be observed only in the hypothyroidism meta-analysis. These disease-specific genetic associations can provide additional insights into the distinct pathophysiology of autoimmune thyroid diseases and can inform the development of disease-specific polygenic risk scores.2. Clinical polygenic risk score (PRS) for thyroid cancer diagnosis.

[0116] This thyroid cancer PRS, defined using independent alleles significantly associated with thyroid cancer in GWAS meta-analysis, achieved an area under the receiver operating characteristic curve (AUC) of 0.686 (95% CI 0.672, 0.7) when identifying individuals diagnosedwith thyroid cancer in the Colorado Center for Personalized Medicine (CCPM) Biobank. The PRS was calculated using effect sizes while excluding CCPM to avoid overfitting (leave-one-biobank-out method). This PRS AUC is significantly better than that developed from the largest publicly available GWAS meta-analysis (GBMI Phase I (22), AUROC 0.652 [0.637, 0.666], DeLong p-value l.le-07), likely due to the greater statistical power of our meta-analysis.

[0117] For the clinically relevant use case of distinguishing ThC from BNG, PRSnic versus BNG is defined as the difference between the PRS for ThC (PRSnic versus AH) and the PRS for BNG (PRSBNG versus AH): PRSThc versus BNG = PRSThc versus AII - PRSBNG versus AII. This approach leverages genetic associations with both ThC and BNG to improve performance for distinguishing benign and malignant thyroid nodules.

[0118] In some implementations, each polygenic risk score is calculated as a weighted sum of genetic variant dosages, wherein each variant dosage represents the number or probability of risk alleles carried by the subject and each weight corresponds to an effect size, regression coefficient, log odds ratio, or other association statistic derived from a genome- wide association study or meta-analysis. For example, a PRS may be calculated according to PRS = X PiGi, where Pi is a variant-specific weight and Gi is the subject’s allele dosage, imputed allele dosage, or genotype call for the corresponding variant. The score may be standardized, normalized, percentile-ranked, decile-ranked, ancestry -calibrated, or adjusted using one or more covariates such as age, sex, genetic principal components, continental ancestry proportions, or clinical variables.

[0119] In some implementations, individuals with PRSnic versus AH in the top decile had 10.7 times the odds of developing ThC than those in the first decile. Thyroid nodules in individuals with PRSrhc versus BNG in the top decile had 7.8 times the odds of being malignant than in individuals with PRSThc versus BNG in the first decile. These odds ratios demonstrate the strong discriminative power of the PRS for identifying individuals at elevated risk of thyroid cancer.

[0120] While a previous study design (thyroid cancer vs. all controls) identifies individuals at risk for thyroid cancer in a population, the present clinical use case is distinct: differentiating thyroid cancers from benign nodules (TC vs. BNG). Electronic chart reviews were performed and identified 1,343 genotyped patients with confirmed diagnoses of non-medullary thyroid cancer and 293 individuals with the diagnosis of benign nodular goiter by surgical histopathology. The earlier thyroid cancer PRS performed suboptimally (AUROC 0.618 [0.581,0.655]) for differentiating thyroid cancers from benign nodular goiters. Therefore, we developed a TC vs. BNG PRS for clinical use by leveraging genetic architecture for both thyroid cancer and benign nodular goiter (FIG. 3, AUC 0.662 [0.626, 0.698]).

[0121] PRS associations with aggressive ThC features were evaluated in three domains (patient, tumor and metastatic disease). PRSTHC versus BNG was significantly associated with the risk of structural disease recurrence, tumor size (T stage), tumor focality, extranodal extension and the number of neck lymph node metastases (Bonferroni-adjusted P < 1.7 x 103). At a nominal P < 0.05, both PRSnic versus AH and PRSihc versus BNG were also associated with death from ThC and locoregional metastases. These associations indicate that the PRS can identify individuals at risk for aggressive, clinically significant thyroid cancer.

[0122] To simulate screening for aggressive ThC (high-risk of structural disease recurrence), PRS performance was tested when all individuals not diagnosed with high-risk ThC, including those diagnosed with low-risk and intermediate-risk ThC, were considered as controls. PRS hc versus AH demonstrated a superior AUC of 0.741 (95% CI 0.682-0.801), sensitivity of 0.803 (0.803-0.803) and specificity of 0.569 (0.565-0.572). The number needed to screen to identify one individual with high-risk ThC in the Colorado Center for Personalized Medicine cohort was 268. For comparison, the US Preventive Services Task Force-recommended screening for colon cancer with colonoscopy has a number needed to screen of 263. Thus, genetically informed screening for high-risk ThC is a conceptually viable strategy to identify aggressive ThCs at an early curable stage to reduce morbidity and mortality.1.2.1. Increase discovery power of thyroid cancer and benign nodular goiter GWAS metaanalyses. The All of Us Research Program Controlled Tier version 8 whole genome sequencing (WGS) data is incorporated into meta-analysis (n=414,000). The UK Biobank WGS genomic data (n=500,000) is reanalyzed to remove the uncertainty associated with imputation and increase variant overlap with other biobanks.1.2.2. Incorporate known pathogenic / likely pathogenic (P / LP) variants for cancer- associated syndromes. Investigation of 245,394 All of Us Research Program participants demonstrated that 1 in 53 thyroid cancer patients carry a known P / LP variant in an actionable cancer-predisposing gene from the American College of Medical Genetics and Genomics list for reporting secondary findings (23). An assessment for known P / LP variants in the PRS is included.Leverage the complete genetic architecture of thyroid disease. A preliminary PRS is constructed from variants significant in GWAS meta-analysis (p-value < 5e-8). Methods that analyze the whole genome (24) tend to outperform the significant hits approach when trained on a large GWAS, such as in the present meta-analysis. Current methods in the field are employed, including LDPred2 (25), MegaPRS (26), PRS-CSx (27), and PRSMix (28).Test PRS performance across ancestries. Ensuring comparable performance in diverse populations is essential for the clinical deployment of a new score. In preliminary analyses, the PRS was tested in two underrepresented populations in the CCPM (AFR and AMR). No significant drop in AUC in AFR and AMR compared to multi-ancestry and EUR populations (FIG. 2B) was found. Rather, the thyroid cancer AUC was significantly higher in AMR individuals (DeLong test p-value < 0.05), likely due to our discovery meta-analysis's high diversity and inclusiveness. The electronic health records are reviewed to add -1000 diverse histologically or cytologically confirmed thyroid cancer and benign multinodular goiter cases, enabling multi-ancestry evaluation of the TC vs. BNG PRS.Multi-ancestry PRS performance. No significant drop in PRS performance measured with AUC was found in the EAS-like, AMR-like and AFR-like strata (DeLong test, P > 0.05) except for hypothyroidism PRS in AFR-like individuals. This demonstrates that the PRS performs equitably across diverse populations, ensuring broad clinical applicability.PRS calibration. To ensure equitable PRS performance in patients of various genetic backgrounds, an adjustment model integrating genetic components of ancestry is fitted to rescale both the mean and variance of the polygenic score within control participants. The residual from this model is used to create ancestry-corrected PRS (29). Alternative modern methods of PRS calibration (30,31) are also tested. Importantly, no single approach works optimally in all circumstances, rather, the best-performing score in our validation cohort (CCPM) is identified, as well as learn about the genetic architecture, portability, and sparsity of genetic associations via which methods perform best in a clinician-refined sample.These steps deliver an optimized PRS for the clinical task of distinguishing benign and malignant thyroid nodules, performing equitably in individuals of EUR, AFR and AMR populations and ready for clinical trial testing (Aim 2). This score is derived from the largest study of thyroid cancer genetics to date, leveraging genetic diversity within studies and across countries and continents.1.3. Ultrasound image-based thyroid nodule classifier.The predictive power of PRS is limited by thyroid cancer's heritability of 28-53% (14,15). Alone, PRS is insufficient for clinical decision-making, which should rely on both genetic and ultrasound-based risk assessments. ACR TI-RADS (3) is a widely used ultrasound-based risk stratification system. Still, it suffers from fair (K = 0.22) interobserver agreement (32,33) and low specificity when evaluated in routine clinical practice (sensitivity 91% and specificity 19%, our unpublished data). Therefore, PRS is combined with the deep learning model that estimates thyroid cancer risk from ultrasound images of thyroid nodules.The current ultrasound-based deep learning thyroid nodule classifier achieved an AUC of 0.83. It is synergistic with the first-generation PRS (16). The accuracy and interpretability of the deep learning model is improved as follows:1.3.1. A much larger training dataset of thyroid nodule ultrasound images for model training is employed. The first-generation model (16) was trained on 32,545 images from 621 nodules. An updated set of 152,300 images and 1534 cine clips from 1516 thyroid nodules is used. The video clip data is leveraged and suitable neural network architectures are tested, such as 3D extensions of convolutional neural network architectures (3D CNN) and newer vision transformer-based methods such as SLIViT (34). This study is inspired by work demonstrating that analyzing 3D nodule shape is helpful even for the most challenging thyroid nodule classification tasks, such as distinguishing follicular thyroid carcinomas from adenomas (35). 1.3.2. A thyroid nodule segmentation model is added to improve prediction accuracy and interpretability. Accordingly, medSAM (36) and medSAM2 (37) models are trained on thyroid nodule still and video ultrasound data, respectively, and segmentation masks are created.1.3.3. The primary deficiency of all published machine learning tools for thyroid cancer diagnosis (including the first-generation model (16)) is that they were trained torecognize two classes, benign and malignant. Thus, these tools perform well in identifying papillary thyroid cancers (PTC) with a suspicious sonographic appearance: a very hypoechoic nodule with microcalcifications. However, these models fail to recognize follicular (FTC) and oncocytic (OTC) thyroid cancers, which frequently lack these suspicious features (38-40). The AUC of our preliminary deep learning model decreased from 0.83 for all thyroid nodules to 0.66 for the difficult- to-diagnose thyroid cancers (unpublished data). To mitigate this deficiency, a deep learning network is trained that recognizes three classes of thyroid nodules: benign, PTC-like cancers and FTC-like cancers. A similar multiclass approach was shown to improve the generalizability and clinical translatability of artificial intelligence-based cervical cancer screening (41,42). The training data is enriched with follicular and oncocytic carcinomas and adenomas. Clinical data and downloaded ultrasound studies have been abstracted for 521 histologically confirmed FTC, OTC, follicular- variant of PTC and benign adenomas. A diffusion network (43,44) is trained to synthesize additional ultrasound images to further augment our training data and improve model performance.

[0123] In summary, the existing ultrasound image-based deep learning thyroid nodule classifier is improved using a larger training image set, taking advantage of video data and training the model to recognize thyroid cancer subtypes.4. Combined genetic and image-based thyroid nodule classifier.

[0124] In some implementations, when the subject’s thyroid-cancer PRS or aggressive-thyroid-cancer PRS satisfies a predefined threshold, the system generates a recommendation for one or more follow-up actions, including thyroid ultrasound screening, earlier imaging, repeat imaging at a shorter interval, endocrinology referral, genetic counseling, molecular diagnostic testing, surgical consultation, active surveillance, thyroid nodule biopsy, or postoperative recurrence monitoring. In some implementations, the threshold is selected to identify subjects at elevated risk for aggressive thyroid cancer while limiting screening of subjects with low-risk or indolent disease. The follow-up recommendation may be generated even before a thyroid nodule is detected, or after a thyroid nodule or biopsy-proven thyroid cancer has been identified.

[0125] The state-of-the-art thyroid nodule classifier uses complementary and synergistic genetic risk assessment (PRS) and image-based risk assessment. The two predictions arecombined using a meta-classifier as previously described (16). The threshold for binary classification (likely malignant - proceed with FNA or likely benign - manage with observation) is defined to achieve a clinically selected operating point, such as > 90% sensitivity and >35%, > 50%>, or greater specificity. Performance is tested on an out-of-sample set of thyroid nodules from genotyped CCPM Biobank participants.

[0126] In some implementations, the meta-classifier receives as inputs one or more of: the composite risk metric, PRS ThC vs. All, PRS BNG vs. All, PRS ThC vs. BNG, an image-based malignancy probability, a benign / PTC-like / FTC-like class probability vector, ultrasound- derived features, segmentation-derived features, TI-RADS features, age, sex, genetic ancestry covariates, clinical history, or other demographic or clinical variables. The meta-classifier may comprise a logistic regression model, linear model, gradient-boosted model, random forest, support vector machine, neural network, Bayesian model, calibrated ensemble model, rules- based model, or weighted combination model. The meta-classifier may output a probability of malignancy, a binary classification, a risk category, a confidence score, or a clinical recommendation. A decision threshold may be selected to achieve a target sensitivity, specificity, negative predictive value, positive predictive value, or reduction in unnecessary FNA procedures.2. A multi-ancestry single-center clinical trial testing the utility of polygenic risk score (PRS) in combination with deep-learning analysis of thyroid ultrasound images for thyroid cancer diagnosis is conducted and validates the (PRS+AI) classifier.

[0127] A blinded single-arm clinical trial is performed to establish the clinical utility of our combined genetic and image-based classifier (PRS+AI) for thyroid nodule diagnosis (FIG. 4). The purpose of the PRS+AI classifier is to improve the triage of thyroid nodules currently evaluated with FNA and to reduce unnecessary procedures for benign thyroid nodules while maintaining clinically acceptable sensitivity for cancer.

[0128] A detailed description of the clinical trial and sample size calculations is provided below. Briefly, patients referred to the University of Colorado are enrolled to evaluate thyroid nodules meeting ACR TI-RADS criteria for FNA biopsy (3). These patients are managed as per current guidelines (43), including FNA, molecular testing and thyroid surgery as indicated. In parallel, blinded investigators assess genetic and image-based thyroid cancer risk using the PRS+AI classifier described herein and FNA or active surveillance is recommended. At theconclusion of the clinical trial, the outcomes of standard-of-care management are abstracted from clinical charts. Recommendations from the PRS+AI classifier are unblinded, and its sensitivity and specificity are calculated.

[0129] Based on our retrospective analysis (16), the PRS+AI achieves a sensitivity > 90%, comparable to that of FNA biopsy (85.6% [95% CI, 79.9-89.5]) (44). The specificity of PRS+AI is expected to be at least 35%. Thus, at least a third of the benign thyroid nodules requiring FNA per the current standard of care can now be recommended for observation. Implementing such a system in the United States reduces the number of FNA for benign thyroid nodules by about 175,000 or more, considering a total of 600,000 FNA procedures per year and a 20% malignancy rate. Consequently, it prevents many inadequate and indeterminate cytology diagnoses, molecular diagnostic tests and diagnostic thyroidectomies for benign thyroid nodules.CLINICAL TRIAL PROTOCOL

[0130] This blinded single-arm clinical trial tests a combined genetic and image-based classifier (PRS+AI) for the triage of thyroid nodules for fine needle aspiration biopsy (FNA) or observation. In this study, the PRS+AI classifier recommendation is generated in parallel with the standard clinical care but does not change clinical management. The 'ground truth' diagnosis of benign or malignant thyroid nodules is available for all clinical trial participants, and the performance of the PRS+AI classifier is evaluated.Inclusion criteria.1. The patient is referred for evaluation of thyroid nodule(s).2. A high-frequency (> 10MHz) ultrasound of the thyroid nodule is performed.3. FNA of a thyroid nodule is indicated per the American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS (3), standard-of-care).4. The patient is willing to undergo thyroid nodule FNA per the recommendation of the treating endocrinologist.5. The patient is willing to undergo thyroid surgery (hemi- or total thyroidectomy) if clinically indicated (FNA cytology Bethesda category IV or V, or Bethesda category III and IV with suspicious molecular diagnostic test (10,11)).6. The patient consents to participate in the study and to undergo a blood draw for DNA extraction and whole genome sequencing.C.2. Exclusion criteria.1. Thyroid nodule FNA with inadequate cytology (FNA cytology Bethesda category I; 'ground truth' diagnosis cannot be established).2. Patients with indeterminate cytology managed with active surveillance ('ground truth' diagnosis cannot be established).3. Patients with a history of bone marrow transplantation (genotypes obtained from blood DNA will not match the patient's inherited genotypes).C.3. Participant enrollment.

[0131] Clinical trial participants are recruited. Many of the -700 new diagnoses of thyroid cancer, e.g., in the state of Colorado (https: / / cancerstatisticscenter.cancer.org / types / thyroid) are made, e.g., at the University of Colorado and UCHealth.

[0132] About 3-4 patients are eligible to enroll in the clinical trial each week. Based on a 50% consent rate and a 10% dropout rate, - 78 patients can consent to the trial each year.

[0133] Only patients with high-quality thyroid ultrasound images available in the picture archiving and communication system are considered for FNA. An endocrinologist specializing in thyroid cancer care reviewed the images to ensure the nodule(s) are eligible for FNA per current clinical guidelines (3,43). All patients seen in the weekly multidisciplinary thyroid cancer clinic receive a comprehensive formal ultrasound evaluation of thyroid and neck lymph nodes on the day of their clinic visit.

[0134] Considering the high incidence of thyroid nodules, the significant number of patients referred to clinics for FNA, policies and protocols ensuring high-quality ultrasound evaluation of every individual and a low risk of participating in the clinical trial, 120 patients are successfully accrued during the trial.

[0135] All eligible adults (> 18 years old), regardless of sex, race, ethnicity, socioeconomic status or health insurance, are recruited into the clinical trial. According to thyroid nodule and thyroid cancer historical data, it is expected that 38% of clinical trial participants are males, 81% of European ancestry, 12% of Hispanic (Admixed American) ancestry and 4% of African American ancestry.C.4. Whole genome sequencing and genetic risk assessment.

[0136] After informed consent is obtained, a blood sample is drawn from the participant and the specimen is transferred to, e.g., the University of Colorado Cancer Center Genomics Corefor DNA extraction (> 500ng of genomic DNA) and whole genome sequencing using rolling circle amplification DNBSEQ-T7 technology (> 3 Ox mean coverage). Quality control metrics (mean and percentage of genome coverage, number of QC30 bases aligned, contamination rate, sex concordance and cross-individual contamination) are used as defined by the All of Us Research Program (45).

[0137] Pseudonymized raw genetic data is processed on the HIPAA-compliant National Institute of Health Terra.bio platform (https: / / terra.bio / ) to protect patient privacy. The Broad Institute's Genome Analysis Toolkit best practices pipeline (46) are followed, including joint genotyping. Polygenic risk score (PRS) is calculated as a weighted sum of the alleles and is calibrated and standardized as described in Section 1.2. Missing genotypes (not to exceed 0.005) are mean-imputed.C.5. Combined genetic and image-based classifier.

[0138] Thyroid ultrasound studies are downloaded from the hospital's picture archiving and communication system in DICOM format. Ultrasound images are reviewed by the research team physician and matched to the clinical notes and images from the FNA procedure. The bounding box outlining the nodule is set for static images (in longitudinal and transverse planes) and video clips. Pixel data corresponding to the nodule are extracted and pseudonymized with the help of our custom software (16). The risk of malignancy is estimated using a trained neural network (Section 1.3). Finally, predictions from PRS and the image-based classifier are combined to produce a recommendation to proceed with FNA or manage thyroid nodules conservatively (Section 1.4). All steps involving PRS, deep learning network and PRS+AI classifier recommendations are performed by a non-clinical member of the research team, who has no access to clinical records and is blinded to the results of clinical thyroid nodule workup.C.6. Outcome.

[0139] All patients are managed per current clinical guidelines (3,43). Physicians review electronic records and record the outcome of thyroid nodule evaluation: benign (thyroid nodules with Bethesda II FNA cytology, indeterminate cytology with low-risk molecular diagnostics, or benign surgical histopathology) or malignant (per surgical histopathology).

[0140] The 'ground truth' diagnosis is available for all or most enrolled patients, allowing calculation of the confusion matrix for the PRS+AI thyroid nodule classifier. The clinical trial has the following outcome:1. PRS+AI specificity of > 35% (that is, >35% benign thyroid nodules are recommended for observation by the PRS+AI system).2. PRS+AI sensitivity (the percentage of malignant thyroid nodules recommended for FNA by the classifier) of > 90%, which is similar to the published sensitivity of the FNA (85.6% [95% CI, 79.9-89.5]) (44). This outcome indicates that clinical use of the PRS+AI classifier is unlikely to increase the number of thyroid cancers missed. While this pilot clinical trial may not possess the statistical power to establish PRS+AI sensitivity, demonstrating that not more than 10% of cancers are missed justifies further testing in a large multi -center clinical trial.BIOSTATISTICAL / BIOINFORMATIC DESIGN AND ANALYSIS

[0141] The present biostatistical / bioinformatic design leverages expertise in statistical genetics, machine learning, and biostatistics from the leaders of the NHLBI-sponsored Population Architecture using Genetics and Epidemiology (PAGE) Study and the NHGRL sponsored Polygenic Risk Methods in Diverse Populations (PRIMED) Consortium, who have developed workshop materials for professional societies on polygenic scoring across diverse populations and approaches for assessing population structure and its impact on complex traits.

[0142] Methods rely heavily on detailed bioinformatics; work is ensured to be interoperable and reproducible. Computer code in R and Python used for the meta-analysis of GWAS, PRS development, and ultrasound image-based deep learning model is publicly available at GitHub (https: / / github.com / npozdey / thyroid_nodule_PRS and https: / / github.com / pozdeyevlab).Clinical trial sample size calculations.

[0143] The clinical trial sample size was calculated using PASS 2022, version 22.0.2 software, based on the following assumptions.1. The prevalence of thyroid cancer among thyroid nodules undergoing FNA biopsy is 20%. This is based on historical data that 23% of FNAs produce malignant or indeterminate cytology (5). Some thyroid nodules with indeterminate cytology are benign. However, the FNA yield is increased by strictly following the TI-RADS FNA criteria that demonstrate higher specificity in a research setting (47).2. The PRS + Al thyroid nodule classifier demonstrates a specificity of at least 50% in a prospective clinical trial. This estimate is based on the preliminary retrospective analysiswhere our first-generation PRS + Al classifier achieved a sensitivity of 95% (95% CI 88-99%) and a specificity of 57% (49-64%) (16).

[0144] The purpose of the PRS+AI classifier is to reduce the number of FNA procedures performed on benign thyroid nodules. We hypothesize that PRS+AI could reduce the number of false positive cases (FNA procedures performed on benign thyroid nodules) by at least 35%; that is, the number of benign nodules recommended for observation identified by PRS+AI is greater than 35% of the total number of benign thyroid nodules recommended for biopsy according to current standard-of-care (ACR TI-RADS (3)).

[0145] A single-group design was used to test whether the proportion of benign cases identified by PRS+AI among the total number of benign thyroid nodules is greater than 0.35 (HO: P < 0.35 versus Hl: P > 0.35). The comparison was made using a one-sided, one-sample exact test, with a target Type I error rate (a) of 0.05. Assuming the proportion of negative cases identified by PRS+AI is 50%, to detect the difference with 90% power, 96 patients with benign thyroid nodules are needed. Considering the 20% malignancy rate among all thyroid nodules undergoing FNA, the total number of patients to be enrolled can be, but not limited to, e.g., 120.SUMMARY

[0146] Evaluating thyroid nodules to rule out malignancy is a common clinical task. Over 600,000 thyroid nodule biopsies are performed annually in the United States. However, only 8% of biopsy results meet cytologic categories requiring thyroid surgery, and most biopsies are potentially avoidable. Improving early assessment of thyroid cancer risk in patients with thyroid nodules is critical to reduce unnecessary procedures, overtreatment and excess costs to the health system.

[0147] Thyroid cancer is highly heritable. Genetic effects contribute up to 53% to thyroid cancer susceptibility. The present invention relates to the concept that thyroid cancer germline genetic assessment is complementary and synergistic to image-based risk evaluation. This is based on a global multi-ancestry meta-analysis of genome- wide association studies aggregating genotype data from 2.9 million individuals.

[0148] As described above, 1. A combined genetic and image-based thyroid nodule classifier (PRS+AI) for thyroid cancer diagnosis is provided. The PRS is optimized for the clinical task of distinguishing benign and malignant thyroid nodules. It is calibrated to performequitably in individuals of diverse ancestry. Alone, PRS is insufficient for clinical decisionmaking, which should rely on both genetic and ultrasound-based risk assessments. Therefore, we improve a first-generation deep learning thyroid nodule classifier using training data of 152,300 images and 1534 cine clips from 1516 thyroid nodules. The combined genetic and image-based meta-classifier is used to identify thyroid nodules as likely benign and recommended for observation or likely malignant and recommended for biopsy.

[0149] Secondly, 2. Patients with thyroid nodules undergo a biopsy to establish a 'ground truth' diagnosis. In parallel, patients are evaluated with PRS+AI. The PRS+AI achieves a sensitivity of > 90%, comparable to that of a biopsy, and a specificity of at least 35%. The result informs a large multi-center clinical trial of PRS+AI and implementation into clinical practice. It is estimated that PRS+AI can reduce the number of procedures for benign thyroid nodules by about 175,000 or more annually in the United States, preventing many unnecessary biopsies and diagnostic thyroidectomies.

[0150] In view of the discovery that the thyroid cancer polygenic risk score (PRS) described herein can identify individuals at risk for aggressive thyroid cancer, applications for the PRS include:

[0151] 1. Population screening for individuals at risk of aggressive thyroid cancer. Rationale: A small but significant proportion of patients with thyroid cancer will develop aggressive, incurable disease. There is currently no curative treatment for these patients, and many will succumb to the disease. PRS can identify patients at risk of aggressive thyroid cancer at an early stage when curative thyroid surgery is possible, thus preventing morbidity and death;

[0152] 2. Improve molecular diagnostic testing for indeterminate thyroid nodules. The existing molecular diagnostic tests rely on detecting somatic mutations and gene expression changes. The PRS described herein provides an alternative, complementary risk assessment based on inherited risk. If incorporated into the existing molecular tests, the PRS improves the performance of the existing molecular tests to provide improved prognostic assessment; and

[0153] 3. Triage of thyroid cancer for surgery vs. active surveillance. Small indolent thyroid cancers can be managed with active surveillance, thus avoiding thyroid surgery. Some thyroid cancers under active surveillance will grow or metastasize to neck lymph nodes. An aggressive thyroid cancer PRS can help identify cancers that can be monitored safely and those that need to be removed surgically. This test, combined with somatic testing, can not only informactive surveillance versus surgery, but can also help guide the extent of surgery (lobectomy versus thyroidectomy) or the choice of thyroid tumor ablation instead of surgery.

[0154] In some implementations, the system is used after thyroid cancer has been diagnosed to inform treatment selection. For example, the PRS, aggressive-thyroid-cancer risk estimate, image-based risk assessment, somatic molecular test result, or combined risk estimate may be used to recommend active surveillance, thermal ablation, lobectomy, total thyroidectomy, lymph-node evaluation, neck dissection, radioactive iodine evaluation, more intensive followup, or standard follow-up. In some implementations, a lower-risk classification supports observation or active surveillance, whereas a higher-risk classification supports surgical removal, a greater extent of surgery, or intensified postoperative monitoring.REFERENCES1. Dean DS, Gharib H. Epidemiology of thyroid nodules. Best Pract Res Clin Endocrinol Metab. 2008;22(6):901-911.2. Durante C, Grani G, Lamartina L, Filetti S, Mandel SJ, Cooper DS. The Diagnosis and Management of Thyroid Nodules: AReview. Jama. 2018;319(9):914-924.3. Tessier FN, Middleton WD, Grant EG, Hoang JK, Berland LL, Teefey SA, Cronan JJ, Beland MD, Desser TS, Frates MC, Hammers LW, Hamper UM, Langer JE, Reading CC, Scoutt LM, Stavros AT. ACR Thyroid Imaging, Reporting and Data System (TI-RADS): White Paper of the ACR TI-RADS Committee. J Am Coll Radiol. 2017; 14(5):587-595.4. Dean DS, Gharib H. Fine-Needle Aspiration Biopsy of the Thyroid Gland. In: Feingold KR, Anawalt B, Boyce A, Chrousos G, Dungan K, Grossman A, Hershman JM, Kaltsas G, Koch C, Kopp P, Korbonits M, McLachlan R, Morley JE, New M, Perreault L, Purnell J, Rebar R, Singer F, Trence DL, Vinik A, Wilson DP, eds. Endotext. South Dartmouth (MA): MDText.com, Inc.; 2015.5. Deaver KE, Haugen BR, Pozdeyev N, Marshall CB. Outcomes of Bethesda categories III and IV thyroid nodules over 5 years and performance of the Afirma gene expression classifier: A single-institution study. Clin Endocrinol (Oxf). 2018;89(2):226-232.6. Nam-Goong IS, Kim HY, Gong G, Lee HK, Hong SJ, Kim WB, Shong YK. Ultrasonography-guided fine-needle aspiration of thyroid incidentaloma: correlation with pathological findings. Clin Endocrinol (Oxf). 2004;60(l):21-28.7. Reuters KB, Mamone M, Tkejiri ES, Camacho CP, Nakabashi CCD, Janovsky C, Yang JH, Andreoni DM, Padovani R, Maciel RMB, Vanderlei FAB, Biscolla RPM. Bethesda Classification and Cytohistological Correlation of Thyroid Nodules in a Brazilian Thyroid Disease Center. Eur Thyroid J. 2018;7(3): 133-138.8. Cawood TJ, Mackay GR, Hunt PJ, O'Shea D, Skehan S, Ma Y. TIRADS Management Guidelines in the Investigation of Thyroid Nodules; Illustrating the Concerns, Costs, and Performance. JEndocr Soc. 2020;4(4):bvaa031.9. Mavromati M, Saiji E, Demarchi MS, Lenoir V, Seipel A, Kuczma P, Jornayvaz FR, Becker M, Fernandez E, De Vito C, Triponez F, Leboulleux S. Unnecessary thyroid surgery rate for suspicious nodule in the absence of molecular testing. Eur Thyroid J. 2023;12(6).10. Alexander EK, Kennedy GC, Baloch ZW, Cibas ES, Chudova D, Diggans J, Friedman L, Kloos RT, LiVolsi VA, Mandel SJ, Raab SS, Rosai J, Steward DL, Walsh PS, Wilde JI, Zeiger MA, Lanman RB, Haugen BR. Preoperative diagnosis of benign thyroid nodules with indeterminate cytology. The New England journal of medicine. 2012;367:705-715.11. Nikiforova MN, Mercurio S, Wald Al, Barbi de Moura M, Callenberg K, Santana-Santos L, Gooding WE, Yip L, Ferris RL, Nikiforov YE. Analytical performance of the ThyroSeq v3 genomic classifier for cancer diagnosis in thyroid nodules. Cancer. 2018; 124(8): 1682- 1690. 12. Nasr CE, Andrioli M, Endo M, Harrell RM, Livhits MJ, Osakwe I, Polavarapu P, Siperstein A, Wei S, Zheng X, Jiang R, Hao Y, Huang JIN, Klopper JP, Kloos RT, Kennedy G, Angell TE. Real-World Performance of the Afirma Genomic Sequencing Classifier (GSC)-A Meta-analysis. J Clin Endocrinol Metab. 2023;108(6):1526-1532.13. Desai D, Lepe M, Baloch ZW, Mandel SJ. ThyroSeq v3 for Bethesda III and IV: An institutional experience. Cancer Cytopathol. 2021; 129(2): 164-170.14. Czene K, Lichtenstein P, Hemminki K. Environmental and heritable causes of cancer among 9.6 million individuals in the Swedish Family-Cancer Database. Int J Cancer.2002;99(2):260-266.15. Lin HT, Liu FC, Lin SF, Kuo CF, Chen YY, Yu HP. Familial Aggregation and Heritability of Nonmedullary Thyroid Cancer in an Asian Population: A Nationwide Cohort Study. J Clin Endocrinol Metab. 2020; 105(7).16. Pozdeyev N, Dighe M, Barrio M, Raeburn C, Smith H, Fisher M, Chavan S, Rafaels N, Shortt JA, Lin M, Leu MG, Clark T, Marshall C, Haugen BR, Subramanian D, Crooks K, GignouxC, Cohen T. Thyroid Cancer Polygenic Risk Score Improves Classification of Thyroid Nodules as Benign or Malignant. J Clin Endocrinol Metab. 2024;109(2):402-412.17. Goldgar DE, Easton DF, Cannon-Albright LA, Skolnick MH. Systematic population-based assessment of cancer risk in first-degree relatives of cancer probands. J Natl Cancer Inst.1994;86(21): 1600-1608.18. Amundadottir LT, Thorvaldsson S, Gudbjartsson DF, Sulem P, Kristjansson K, Amason S, Gulcher JR, Bjomsson J, Kong A, Thorsteinsdottir U, Stefansson K. Cancer as a complex phenotype: pattern of cancer distribution within and beyond the nuclear family. PLoS Med.2004; l(3):e65.19. Pozdeyev N, Gay L, Sokol ES, Hartmaier RJ, Deaver KE, Davis SN, French JD, Vanden Borre P, LaBarbera DV, Tan AC, Schweppe RE, Fishbein L, Ross JS, Haugen BR, Bowles DW. Genetic analysis of 779 advanced differentiated and anaplastic thyroid cancers. Clin Cancer Res.2018.20. Ibrahimpasic T, Xu B, Landa I, Dogan S, Middha S, Seshan V, Deraje S, Carlson DL, Migliacci J, Knauf JA, Untch B, Berger MF, Morris L, Tuttle RM, Chan T, Fagin JA, Ghossein R, Ganly I. Genomic Alterations in Fatal Forms of Non-Anaplastic Thyroid Cancer: Identification of MED 12 and RBM10 as Novel Thyroid Cancer Genes Associated with Tumor Virulence. Clin Cancer Res. 2017;23(19):5970-5980.21. Martin AR, Kanai M, Kamatani Y, Okada Y, Neale BM, Daly MJ. Clinical use of current polygenic risk scores may exacerbate health disparities. Nat Genet. 2019;51(4):584-591.22. Zhou W, Kanai M, Wu KH, Rasheed H, T suo K, Hirbo JB, Wang Y, Bhattacharya A, Zhao H, Namba S, Surakka I, Wolford BN, Lo Faro V, Lopera-Maya EA, Lail K, Fave MJ, Partanen JJ, Chapman SB, Karjalainen J, Kurki M, Maasha M, Brumpton BM, Chavan S, Chen TT, Daya M, Ding Y, Feng YA, Guare LA, Gignoux CR, Graham SE, Hornsby WE, Ingold N, Ismail SI, Johnson R, Laisk T, Lin K, Lv J, Millwood IY, Moreno-Grau S, Nam K, Palta P, Pandit A, Preuss MH, Saad C, Setia-Verma S, Thorsteinsdottir U, Uzunovic J, Verma A, Zawistowski M, Zhong X, Afifi N, Al-Dabhani KM, Al Thani A, Bradford Y, Campbell A, Crooks K, de Bock GH, Damrauer SM, Douville NJ, Finer S, Fritsche LG, Fthenou E, Gonzalez-Arroyo G, Griffiths CJ, Guo Y, Hunt KA, loannidis A, Jansonius NM, Konuma T, Lee MTM, Lopez -Pineda A, Matsuda Y, Marioni RE, Moatamed B, Nava-Aguilar MA, Numakura K, Patil S, Rafaels N, Richmond A, Rojas-Munoz A, Shortt JA, Straub P, Tao R, Vanderwerff B, Vemekar M, Veturi Y, Barnes KC,Boezen M, Chen Z, Chen CY, Cho J, Smith GD, Finucane HK, Franke L, Gamazon ER, Ganna A, Gaunt TR, Ge T, Huang H, Huffman J, Katsanis N, Koskela JT, Lajonchere C, Law MH, Li L, Lindgren CM, Loos RJF, MacGregor S, Matsuda K, Olsen CM, Porteous DJ, Shavit JA, Snieder H, Takano T, Trembath RC, Vonk JM, Whiteman DC, Wicks SJ, Wijmenga C, Wright J, Zheng J, Zhou X, AwadallaP, Boehnke M, Bustamante CD, Cox NJ, Fatumo S, Geschwind DH, Hayward C, Hveem K, Kenny EE, Lee S, Lin YF, Mbarek H, Magi R, Martin HC, Medland SE, Okada Y, Palotie AV, Pasaniuc B, Rader DJ, Ritchie MD, Sanna S, Smoller JW, Stefansson K, van Heel DA, Walters RG, Zbllner S, Martin AR, Wilier CJ, Daly MJ, Neale BM. Global Biobank Metaanalysis Initiative: Powering genetic discovery across human disease. Cell Genom.2022;2(10): 100192.23. Green RC, Berg JS, Grody WW, Kalia SS, Korf BR, Martin CL, McGuire AL, Nussbaum RL, O'Daniel JM, Ormond KE, Rehm HL, Watson MS, Williams MS, Biesecker LG. ACMG recommendations for reporting of incidental findings in clinical exome and genome sequencing. Genet Med. 2013;15(7):565-574.24. Ni G, Zeng J, Revez JA, Wang Y, Zheng Z, Ge T, Restuadi R, Kiewa J, Nyholt DR, Coleman JRI, Smoller JW, Yang J, Visscher PM, Wray NR. A Comparison of Ten Polygenic Score Methods for Psychiatric Disorders Applied Across Multiple Cohorts. Biol Psychiatry.2021;90(9):611-620.25. Prive F, Arbel J, Vilhjalmsson BJ. LDpred2: better, faster, stronger. Bioinformatics.2020;36(22-23):5424-5431.26. Zhang Q, Prive F, Vilhjalmsson B, Speed D. Improved genetic prediction of complex traits from individual-level data or summary statistics. Nat Commun. 2021 ; 12( 1):4192.27. Ruan Y, Lin YF, Feng YA, Chen CY, Lam M, Guo Z, He L, Sawa A, Martin AR, Qin S, Huang H, Ge T. Improving polygenic prediction in ancestrally diverse populations. Nat Genet.2022;54(5):573-580.28. Truong B, Hull LE, Ruan Y, Huang QQ, Hornsby W, Martin H, van Heel DA, Wang Y, Martin AR, Lee SH, Natarajan P. Integrative polygenic risk score improves the prediction accuracy of complex traits and diseases. Cell Genom. 2024;4(4): 100523.29. Khera AV, Chaffin M, Zekavat SM, Collins RL, Roselli C, Natarajan P, Lichtman JH, D'Onofrio G, Mattera J, Dreyer R, Spertus JA, Taylor KD, Psaty BM, Rich SS, Post W, Gupta N, Gabriel S, Lander E, Ida Chen YD, Talkowski ME, Rotter JI, Krumholz HM, Kathiresan S. Whole-Genome Sequencing to Characterize Monogenic and Polygenic Contributions in Patients Hospitalized With Early-Onset Myocardial Infarction. Circulation. 2019; 139(13): 1593-1602. 30. Ge T, Irvin MR, Patki A, Srinivasasainagendra V, Lin YF, Tiwari HK, Armstrong ND, Benoit B, Chen CY, Choi KW, Cimino JJ, Davis BH, Dikilitas O, Etheridge B, Feng YA, Gainer V, Huang H, Jarvik GP, Kachulis C, Kenny EE, Khan A, Kiryluk K, Kottyan L, Kullo IJ, Lange C, Lennon N, Leong A, Malolepsza E, Miles AD, Murphy S, Namjou B, Narayan R, O'Connor MJ, Pacheco JA, Perez E, Rasmussen-Torvik LJ, Rosenthal EA, Schaid D, Stamou M, Udler MS, Wei WQ, Weiss ST, Ng MCY, Smoller JW, Lebo MS, Meigs JB, Limdi NA, Karlson EW. Development and validation of a trans-ancestry polygenic risk score for type 2 diabetes in diverse populations. Genome Med. 2022;14(l):70.31. Lennon NJ, Kottyan LC, Kachulis C, Abul-Husn NS, Arias J, Belbin G, Below JE, Berndt SI, Chung WK, Cimino JJ, Clayton EW, Connolly JJ, Crosslin DR, Dikilitas O, Velez Edwards DR, Feng Q, Fisher M, Freimuth RR, Ge T, Glessner JT, Gordon AS, Patterson C, Hakonarson H, Harden M, Harr M, Hirschhom JN, Hoggart C, Hsu L, Irvin MR, Jarvik GP, Karlson EW, Khan A, Khera A, Kiryluk K, Kullo I, Larkin K, Limdi N, Linder JE, Loos RTF, Luo Y, Malolepsza E, Manolio TA, Martin LJ, McCarthy L, McNally EM, Meigs JB, Mersha TB, Mosley JD, Musick A, Namjou B, Pai N, Pesce LL, Peters U, Peterson JF, Prows CA, Puckelwartz MJ, Rehm HL, Roden DM, Rosenthal EA, Rowley R, Sawicki KT, Schaid DJ, Smit RAJ, Smith JL, Smoller JW, Thomas M, Tiwari H, Toledo DM, Vaitinadin NS, Veenstra D, Walunas TL, Wang Z, Wei WQ, Weng C, Wiesner GL, Yin X, Kenny EE. Selection, optimization and validation of ten chronic disease polygenic risk scores for clinical implementation in diverse US populations. Nat Med.2024;30(2):480-487.32. Hoang JK, Middleton WD, Farjat AE, Teefey SA, Abinanti N, Boschini FJ, Bronner AJ, Dahiya N, Hertzberg BS, Newman JR, Scanga D, Vogler RC, Tessier FN. Interobserver Variability of Sonographic Features Used in the American College of Radiology Thyroid Imaging Reporting and Data System. AJR Am J Roentgenol. 2018;211(1): 162-167.33. Lee HJ, Yoon DY, Seo YL, Kim JH, Baek S, Lim KJ, Cho YK, Yun EJ. Intraobserver and Interobserver Variability in Ultrasound Measurements of Thyroid Nodules. J Ultrasound Med.2018;37(l): 173-178.34. Avram O, Durmus B, Rakocz N, Corradetti G, An U, Nitalla MG, Rudas A, Wakatsuki Y, Hirabayashi K, Velaga S, Tiosano L, Corvi F, Verma A, Karamat A, Lindenberg S, Oncel D,Almidani L, Hull V, Fasih-Ahmad S, Esmaeilkhanian H, Wykoff CC, Rahmani E, Arnold CW, Zhou B, Zaitlen N, Gronau I, Sankararaman S, Chiang JN, Sadda SR, Halperin E. SLIViT: a general Al framework for clinical-feature diagnosis from limited 3D biomedical-imaging data. Res Sq. 2023.35. Bell C, White SL, Tylee T, Dighe M, Greca A, Goldner W, May son S, Haugen BR, Pozdeyev N. Thyroid Nodule Sphericity Metrics Discriminate Benign and Malignant Follicular and Oncocytic Neoplasms. Thyroid. 2025;35(3):291-297.36. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203-211.37. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. Paper presented at: 2017 IEEE International Conference on Computer Vision (ICCV); 22-29 Oct. 2017, 201738. Matrone A, Gambale C, Pieroni E, De Napoli L, Torregrossa L, Materazzi G, Elisei R. Ultrasound features and risk stratification system in NIFT-P and other follicular-patterned thyroid tumors. Eur J Endocrinol. 2023;189(2):175-182.39. Sillery JC, Reading CC, Charboneau JW, Henrichsen TL, Hay ID, Mandrekar JN. Thyroid follicular carcinoma: sonographic features of 50 cases. AJR Am J Roentgenol. 2010;194(l):44-54.40. Maizlin ZV, Wiseman SM, Vora P, Kirby JM, Mason AC, Filipenko D, Brown JA. Hurthle cell neoplasms of the thyroid: sonographic appearance and histologic characteristics. J Ultrasound Med. 2008;27(5):751-757; quiz 759.41. Ho J, Jain A, Abbeel P. Denoising Diffusion Probabilistic Models.2020 :arXiv: 2006.11239. https: / / ui.adsabs.harvard.edu / abs / 2020arXiv200611239H. Accessed June 01, 202042. Cao H, Tan C, Gao Z, Xu Y, Chen G, Heng P-A, Li SZ. A Survey on Generative Diffusion Model. 2022 :arXiv: 2209.02646. https: / / ui.adsabs.harvard.edu / abs / 2022arXiv220902646C. Accessed September 01, 202243. Haugen BR, Alexander EK, Bible KC, Doherty GM, Mandel SJ, Nikiforov YE, Pacini F, Randolph GW, Sawka AM, Schlumberger M, Schuff KG, Sherman SI, Sosa JA, Steward DL, Tuttle RM, Wartofsky L. 2015 American Thyroid Association Management Guidelines for AdultPatients with Thyroid Nodules and Differentiated Thyroid Cancer: The American Thyroid Association Guidelines Task Force on Thyroid Nodules and Differentiated Thyroid Cancer. Thyroid. 2016;26(l): 1-133.44. Hsiao V, Massoud E, Jensen C, Zhang Y, Hanlon BM, Hitchcock M, Arroyo N, Chiu AS, Fernandes-Taylor S, Alagoz O, Sundling K, LiVolsi V, Francis DO. Diagnostic Accuracy of Fine-Needle Biopsy in the Detection of Thyroid Malignancy: A Systematic Review and Meta-analysis. JAMA Surg. 2022; 157(12): 1105-1113.45. Genomic data in the All of Us Research Program. Nature. 2024;627(8003):340-346. 46. van der Auwera G, O'Connor BD. Genomics in the Cloud: Using Docker, GATK, and WDL in Terra. O'Reilly Media, Incorporated.47. Li W, Wang Y, Wen J, Zhang L, Sun Y. Diagnostic Performance of American College of Radiology TLRADS: A Systematic Review and Meta- Analysis. AJR Am J Roentgenol.2021;216(l):38-47.48. White, S.L., Brasher, M.S., Pattee, J. et al. Global multi-ancestry genome-wide analyses identify genes and biological pathways associated with thyroid cancer and benign thyroid diseases. Nat Genet 58, 307-316 (2026). https: / / doi.org / 10.1038 / s41588-025-02483-w.All references, publications, etc. mentioned are incorporated by reference herein in their entireties for their teachings.

Claims

THAT WHICH IS CLAIMED:

1. A computer-implemented method for classifying a thyroid nodule in a subject, the method comprising:receiving genetic variant data associated with the subject;determining, based on the genetic variant data,a first polygenic risk score associated with thyroid cancer, anda second polygenic risk score associated with benign thyroid nodular disease; generating a composite risk metric based on a combination of the first polygenic risk score and the second polygenic risk score, wherein the combination reflects differential genetic contributions to malignant and benign thyroid nodular conditions;classifying the thyroid nodule as malignant or benign based on the composite risk metric; andgenerating an output indicative of the classification for use in clinical decision-making.

2. The method of claim 1, wherein the composite risk metric comprises a difference between the first polygenic risk score and the second polygenic risk score.

3. The method of claim 1, wherein the composite risk metric comprises a weighted combination of the first polygenic risk score and the second polygenic risk score.

4. The method of claim 1, wherein the first polygenic risk score and the second polygenic risk score are derived from genetic variants identified through multi-ancestry genome-wide association studies.

5. The method of claim 1, wherein the first polygenic risk score is derived from variants associated with malignant thyroid disease and the second polygenic risk score is derived from variants associated with benign thyroid nodular disease.

6. The method of claim 1, further comprising:receiving ultrasound image data of the thyroid nodule;processing the ultrasound image data using a trained deep learning model to generate an image-based risk assessment; andcombining the image-based risk assessment with the composite risk metric to generate a combined classification.

7. The method of claim 6, wherein the trained deep learning model is configured to classify thyroid nodules into at least three classes comprising benign, papillary thyroid cancer-like, and follicular thyroid cancer-like.

8. The method of claim 6, wherein the ultrasound image data comprises one or more of static images and video clips of the thyroid nodule.

9. The method of claim 1, wherein the output comprises a recommendation regarding biopsy, surveillance, or surgical intervention.

10. A computer-implemented method for identifying a subject at risk for aggressive thyroid cancer, the method comprising:receiving germline genetic data of the subject;calculating a polygenic risk score for thyroid cancer based on a plurality of genetic variants associated with the subject;determining an association between the polygenic risk score and one or more aggressive thyroid cancer phenotypes;estimating a risk of aggressive thyroid cancer for the subject based on the polygenic risk score; andgenerating an output indicative of the estimated risk for use in clinical screening, imagining, or intervention.

11. The method of claim 10, wherein the one or more aggressive thyroid cancer phenotypes comprise one or more of: tumor size, tumor focality, lymph node metastasis, extranodal extension, structural disease recurrence, or death from thyroid cancer.

12. The method of claim 10, wherein the subject is classified as a high risk when the polygenic risk score exceeds a predefined threshold.

13. The method of claim 12, wherein the predefined threshold corresponds to a top decile of a polygenic risk score distribution.

14. The method of claim 10, wherein the polygenic risk score is derived from genome-wide significant variants identified in a multi -ancestry meta-analysis comprising data from a plurality of biobanks across multiple countries and genetic ancestries.

15. The method of claim 10, wherein the method is used for population-level screening to identify individuals at elevated risk of aggressive thyroid cancer.

16. The method of claim 10, wherein the output comprises a recommendation for targeted screening, earlier imaging, or preventative intervention.

17. The method of any one of claims 10-16, further comprising:receiving ultrasound image data of a thyroid nodule of the subject;processing the ultrasound image data using a trained deep learning model to generate an image-based risk assessment; andcombining the image-based risk assessment with the polygenic risk score to generate a combined risk estimate for the subject.

18. A system for classifying a thyroid nodule, the system comprising:one or more processors; anda memory storing instructions that, when executed by the one or more processors, cause the system to:receive genetic variant data associated with a subject;determine a first polygenic risk score associated with thyroid cancer and a second polygenic risk score associated with benign thyroid nodular disease;generate a composite risk metric based on a combination of the first polygenic risk score and the second polygenic risk score;receive ultrasound image data of a thyroid nodule of the subject;process the ultrasound image data using a trained deep learning model to generate an image-based risk assessment;combine the composite risk metric and the image-based risk assessment to classify the thyroid nodule as malignant or benign; andgenerate an output indicative of the classification.

19. The system of claim 18, wherein the composite risk metric comprises a difference between the first polygenic risk score and the second polygenic risk score.

20. The system of claim 18, wherein the trained deep learning model is configured to classify thyroid nodules into at least three classes comprising benign, papillary thyroid cancer-like, and follicular thyroid cancer-like.

21. The system of claim 18, wherein the first polygenic risk score and the second polygenic risk score are calibrated to provide equitable risk assessment across subjects of diverse genetic ancestries.

22. The system of claim 18, wherein the ultrasound image data comprises video clips and the trained deep learning model comprises a three-dimensional convolutional neural network or a vision transformer architecture.

23. The system of claim 18, wherein the instructions further cause the system to generate a recommendation to proceed with fine needle aspiration biopsy or to manage the thyroid nodule with observation based on the classification.

24. The system of claim 18, wherein the classification achieves a sensitivity of at least 90% and a specificity of at least 35% for detecting thyroid cancer.

25. The system of any one of claims 18-24, wherein the first polygenic risk score and the second polygenic risk score are derived from genetic variants identified through a meta-analysis of genome-wide association studies aggregating data from a plurality of biobanks across multiple countries and genetic ancestries.