Bilinear Pooling for Thyroid Nodule Classification in Ultrasound Imaging

Thyroid cancer is a type of cancer that affects the thyroid gland. It appears as a nodule. The most used imaging modality for thyroid diagnosis is the ultrasound technique. This technique differentiates between thyroid cancer types (benign or malignant) at early-stage. However, visual diagnosis of thyroid nodules in ultrasound images is influenced by physicians' experience. Consequently, using an automatic thyroid diagnosis tool can help physicians in their image interpretation and improve the objectivity and accuracy of thyroid nodule analysis. Recently, deep learning-based algorithms allow an accurate and efficient classification performance for thyroid nodule classification in ultrasound images. Therefore, in the present study, we propose a novel architecture based on the bilinear convolutional concept that consists of a fusion of the outputs of two CNN models using outer products for thyroid nodule classification in ultrasound images. Eleven convolutional neural networks (CNN) and bilinear convolutional neural network (BCNN) algorithms, including VGG-16, VGG-19, ResNet-50, Inception-V3, BCNN(InceptionV3)(2), BCNN(VGG-16)(2), BCNN(VGG-19)(2), BCNN(ResNet-50)(2), BCNN(VGG-19,VGG-16), BCNN(VGG-16,ResNet-50), and BCNN(VGG-19,ResNet-50), are applied in this paper. The proposed approach is evaluated on a public dataset that contains 447 ultrasound images of thyroid nodules. Findings show that BCNN algorithms outperformed CNN architectures in thyroid nodule classification in ultrasound images. The proposed approach can be used to support the interpretation and analysis of thyroid nodules in ultrasound images. It is a second opinion that can reduce morbidity and improve the thyroid nodule classification.