Benign and Malignant Classification of Thyroid Nodules Based on Bilinear Convolutional Neural Network

Thyroid cancer is a malignant tumor occurring in thyroid tissues, posing a significant threat to human health. Ultrasonography is the most commonly used imaging modality for thyroid nodules and is widely employed in their diagnostic evaluation. With the advancement of deep learning, automated diagnostic tools based on ultrasound images have emerged to assist clinicians in classifying thyroid nodules as benign or malignant, thereby enhancing diagnostic accuracy. In this paper, we propose a bilinear convolutional neural network model named THBCNN. The architecture incorporates two parallel feature extraction subnetworks: a custom-designed network termed SubThyroidNet, and an improved version of ResNet34. The feature maps output by the two sub-networks are fused through bilinear pooling, enabling the model to make more accurate judgments on the benign and malignant nature of thyroid nodules across multiple feature dimensions. Experimental results show that the accuracy of THBCNN in identifying benign and malignant thyroid nodules reaches 0.9576, which is a significant improvement compared with common classification models.