Analyzing medical images has always been a time-consuming and challenging task, as this requires massive precision and experienced medical specialists to handle. Over the past few years, deep learning (DL) has remarkably developed to help address that demand, one of its application in healthcare is diagnosing thyroid nodules from ultrasound images. The key factor for any DL model to function properly is that they are trained with a sufficient database, which is hard to achieve since medical images are difficult to acquire in a big number. In this paper, we propose a pretrained Inception-v3 model integrated with Grid Search Optimization (GSO) technique to determine whether a thyroid nodule in an ultrasound image is benign or malignant. We also operated various data augmentation techniques, and with the presence of Transfer Learning (TL), the problem of limited amount of data would somewhat be solved. Using four evaluation metrics: Accuracy, Specificity, Sensitivity and F-measure, we obtained very encouraging results.