Transfer Learning in Convolutional Neural Network to Differentiate Follicular Adenoma Versus Follicular Carcinoma of Thyroid on Aspiration Cytology Material

AIMS AND OBJECTIVES: In the present paper, we tried to build a basic convolutional neural network model and a transfer learning model on it to distinguish follicular adenoma (FA) and follicular carcinoma (FC) of thyroid in fine needle aspiration cytology (FNAC). MATERIALS AND METHODS: We selected histopathology proven cases of FA (25 cases) and FC (29 cases). In each case, the best May Grunwald Giemsa stained smear was selected and 10 to 11 representative microphotographs were taken by Olympus microscope (Olympus DP74) in 40× objectives from the most representative area of the smear. There were total 347 microphotographs in FA and 229 microphotographs of FC. The images were divided into training (406 images, 70%), validation (89 images, 15%) and test set (81 images, 15%). The basic (scratch) model of convolutional neural network (CNN) was built in Python 3.11.11. We subsequently, also used a transfer learning model by a pre-trained MobileNetV2 in the same images. We ran the model in Jupyter notebook for 15 epochs with 13 steps in each epoch. RESULT: The sensitivity and specificity of the base model were 72.41% and 88.46%, respectively. The accuracy, precision and F1 score were 82.71%, 77.78% and 75.00%. The area under the curve (AUC) of receiver operating characteristic (ROC) was 0.85. The sensitivity and specificity of the transfer learning model are 82.76% and 92.31%, respectively. The accuracy, precision and F1 score were 88.89%, 85.71% and 84.21%. The area under the curve (AUC) of receiver operating characteristic (ROC) is 0.94. CONCLUSIONS: We built a CNN base model and transfer learning model (MobileNetV2) and successfully distinguished FA and FC in thyroid FNAC. To the best of our knowledge, this is the first study of CNN in thyroid follicular tumours.