Computer-Aided Diagnosis of Thyroid Nodule from Ultrasound Images Using Transfer Learning from Deep Convolutional Neural Network Models

Nowadays, thyroid cancer is considered as one of the most common endocrine cancer in the human body. Ultrasonography is the primary imaging modality for the diagnosis of thyroid cancer. Computer-Aided assessment of ultrasound images for differentiating malignant nodules from benign nodule may help the clinicians for their decision making, and it leads to early diagnosis and on-time treatment. The important problem is difficulty in capturing features appropriate for differentiating malignant nodules from benign nodules. In this study, we extensively investigated the feasibility of transfer learning technique for the extraction of high-level features from thyroid ultrasound images. Images are preprocessed to adjust the skewed distribution using a cluster-based sampling technique. Pre-trained convolutional neural network models are fine-tuned with these preprocessed Images for the extraction of high-level semantic features from Images. Then the extracted features are fed into several supervised learning algorithms, and the performance of each model is evaluated. The experimental results recommend the viability of the Inception-v3 network and Xception network for efficiently differentiating malignant thyroid nodules from benign nodules.