Abstract:Abstract Background Preoperative diagnosis of thyroid lesions by fine-needle aspiration (FNA) biopsy cytology can be challenging, and in up to 20% of cases is unachievable. Patients with an indeterminate preoperative diagnosis are often recommended for diagnostic surgery, with the majority receiving a benign diagnosis, rendering the surgery unnecessary. Consequently, new technologies that can provide accurate preoperative diagnosis of thyroid lesions are needed. Here, we have employed DESI-MS imaging along with statistical modeling to determine molecular signatures of benign and malignant thyroid lesions using banked thyroid tissue samples with known histopathology diagnosis. We then applied this methodology for analysis and classification of preoperatively collected FNA smears. Methods A total of 199 fresh-frozen thyroid tissues including 50 normal, 55 follicular adenoma (FTA), 58 follicular carcinoma (FTC), and 36 papillary carcinoma (PTC) were sectioned at a thickness of 10 μm and stored at −80 °C until analysis. FNA biopsies were prospectively collected at Baylor College of Medicine. DESI-MS imaging was performed in the negative ion mode using a Waters Xevo G2-XS mass spectrometer fitted with a DESI-XS source. Samples were H&E stained and pathologically evaluated after analysis. Molecular profiles from tissue regions of clear histology were used to build classification models. For the FNA smears, mass spectra corresponding to clusters of thyroid cells were extracted for statistical prediction, and the predictive performance of the models on FNA smears was assessed in correlation with pathology. Results DESI-MS analysis of thyroid tissue sections generated 151 317 individual mass spectra, comprised of hundreds of lipid and metabolite features, that were used to build and validate two classification models: PTC vs benign thyroid, comprised of normal thyroid and FTA, and FTC vs benign thyroid. For each of these models, the data was randomly split with two-thirds used as a training set to generate the model and one-third used as an independent validation set to assess the performance of the model. For the PTC vs benign thyroid model, a prediction accuracy of 97.7% was achieved for the training set with an accuracy of 96.6% for the withheld validation dataset. For the FTC vs benign model, an accuracy of 78.4% was achieved for both the training and validation datasets. We are currently refining statistical workflows to improve the performance of the FTC model and to generate a model discriminating benign thyroid from thyroid cancer (FTC and PTC combined). When testing the prediction of the classifiers built from tissue imaging data on an initial sample set of 19 FNAs, overall per-sample prediction accuracies of 79% and 89% were obtained using the PTC vs benign thyroid model and the FTC vs benign thyroid model, respectively. We are currently working to collect, analyze, and validate the performance of the classification models on additional FNA biopsies, with a focus on samples with indeterminate cytology. Conclusions Overall prediction accuracies of 79% and 89% were achieved for the benign vs PTC model and benign vs FTC model, respectively, for thyroid FNA classification. With the addition of DESI-MS, unnecessary diagnostic surgeries could be prevented by providing improved preoperative specificity.

Use of natural language processing to extract and classify papillary thyroid cancer features from surgical pathology reports

Use of natural language processing to extract and classify papillary thyroid cancer features from surgical pathology reports

A Systematic Review of Natural Language Processing Methods and Applications in Thyroidology

Ambiguous and Incomplete: Natural Language Processing Reveals Problematic Reporting Styles in Thyroid Ultrasound Reports

Identification of pancreatic cancer risk factors from clinical notes using natural language processing

Natural Language Processing to Identify Abnormal Breast, Lung, and Cervical Cancer Screening Test Results from Unstructured Reports to Support Timely Follow-up.

Extracting Thyroid Nodules Characteristics from Ultrasound Reports Using Transformer-based Natural Language Processing Methods

Risk stratification of papillary thyroid cancers using multidimensional machine learning

An accessible, efficient, and accurate natural language processing method for extracting diagnostic data from pathology reports

Protein Classifier for Thyroid Nodules Learned from Rapidly Acquired Proteotypes

Novel approach to implementing natural language processing for clinical staging of non-small-cell lung cancer.

Development and Validation of a Natural Language Processing Algorithm for Extracting Clinical and Pathological Features of Breast Cancer From Pathology Reports

Natural language processing for automated breast cancer recurrence detection and classification in computed tomography reports.

AI-Based multimodal Multi-tasks analysis reveals tumor molecular heterogeneity, predicts preoperative lymph node metastasis and prognosis in papillary thyroid carcinoma: A retrospective study

Highly accurate diagnosis of papillary thyroid carcinomas based on personalized pathways coupled with machine learning

How Natural Language Processing Can Aid With Pulmonary Oncology Tumor Node Metastasis Staging From Free-Text Radiology Reports: Algorithm Development and Validation

Risk Factors and Diagnostic Prediction Models for Papillary Thyroid Carcinoma.

Development of a privacy preserving large language model for automated data extraction from thyroid cancer pathology reports

Development of a Structured Query Language and Natural Language Processing Algorithm to Identify Lung Nodules in a Cancer Centre

Thyroid cancer polygenic risk score combined with deep learning analysis of ultrasound images improves the classification of thyroid nodules as benign or malignant

B-200 Preoperative Classification of Thyroid Nodules by Desorption Electrospray Ionization Mass Spectrometry Imaging of Fine Needle Aspiration Biopsies