Frozen section examination could provide pathological diagnosis for surgery of thyroid nodules, which is time-consuming, skill- and experience-dependent. This study developed a rapid classification method for thyroid nodules and machine learning. Total 69 tissues were collected including 43 nodules and 26 nodule-adjacent tissues. Intraoperative frozen section was first performed to give accurate diagnosis, and the rest frozen specimen were pretreated for probe electrospray ionization mass measurement. By multivariate analysis of mass scan data, a series compounds were found downregulated in the extraction solution of papillary thyroid carcinoma (PTC), but some were found upregulated by mass spectrometry imaging. m/z 758.5713 ([PC[34:2] + H](+)), m/z 772.5845 ([PC[32:0] + K](+)), and m/z 786.6037 ([PC[36:2] + H](+)) were firstly identified as potential biomarkers for nodular goiter (NG). Machine learning was employed by means of support vector machine (SVM) and random forest (RF) algorithms. For classification of PTC from NG, SVM and RF algorithms exhibited the same performance and the concordance was 94.2% and 94.4% between prediction and pathological diagnosis with positive and negative mass dataset, respectively. For the classification of PTC from PTC adjacent tissues, SVM was better than RF and the concordance was 93.8% and 83.3% with positive and negative mass dataset, respectively. With the identified compounds as training features, the sensitivity and specificity are 87.5% and 88.9% for the test set. The developed method could also correctly predict the malignancy of one medullary thyroid carcinoma and one adenomatous goiter (benign). The diagnosis time is about 10 min for one specimen, and it is very promising for the intraoperative diagnosis of papillary thyroid carcinoma.