Thyroid Ultrasonography Image Classification Based on Fine-tuned Convolutional Neural Network

Abstract

Most commonly found thyroid nodules are benign which is less harmful in comparison to malignant nodules. Number of techniques are available such as Ultrasonography imaging, percutaneous biopsy to determine whether a nodule is benign or malignant. However, these techniques require well experienced and senior radiologists. Only benignity and malignancy classification sometime result unnecessary surgery. Current Classification scheme, Thyroid Imaging Reporting and Data System (TIRADS) further classified the benign and malignant nodule which preclude biopsies required or not. The ensemble RetinaNet in conjunction with US image which improve nodule characterization and reduce biopsies. RetinaNet is promising technique as it is a simpler one-stage object detector which is fast and efficient. RetinaNet has been proven to perform conventional object detection tasks but has not been tested on detecting in Thyroid nodules. Here ensemble RetinaNet has been implemented which classified thyroid nodules based on TIRADS classes successfully. To validate its performance, the experimental setup has been constructed using the thyroid digital image database (TDID). In addition to training and testing on the same dataset, evaluation of model set up is done by pre-trained ImageNet dataset. The diagnostic performance of the ensemble network model was calculated on the basis of precision, recall and F1 value. The precision value of the aforementioned network obtained up to 94% while recall value obtained up to 96% and F1 score obtained up to 93%.