Application of machine learning algorithm for cytological diagnosis of thyroid cancer

The machine learning algorithm (MLA) is a promising tool for cytological diagnosis of thyroid cancer. We developed the MLA classifying human thyroid cell clusters using brightfield microscope-based color images and refractive index information-based (RI) images and evaluated the complementarity of the two types of images. Color and RI images of thyroid fine-needle aspiration biopsy (FNAB) specimen were taken using brightfield microscope and Optical Diffraction Tomography. MLA was designed to classify benign and malignant cell clusters using color images, RI images, or both. This study included 1295 thyroid cell clusters (benign: malignancy = 964: 331) obtained from 110 patients. The accuracies of MLA-classifiers using color images, RI images, and both were 99.0%, 97.5%, and 100%, respectively. In the model using color image, the correlation between the size of the nuclei and the prediction score for the probability of cancer was prominent. In contrast, the model using RI image showed high confidence when the degree of detail of the images surrounding the nuclei quantified using the Brenner gradient was high. Qualitative assessment using gradient-weighted class activation mapping and T-distributed stochastic neighbor embedding analysis also showed that the size of the nucleus was mainly used in the color image, but detailed morphological information of the nucleus was used in the RI image. MLA has great potential for diagnosing thyroid cancer based on imaging of FNAB specimens. Complementary information from color images and RI images can improve the performance of MLA.