Thyroid Nodule Benignty Prediction by Deep Feature Extraction

Thyroid nodules are a common pathology which are fortunately usually benign. However, current image characterization is limited in accurately differentiating benign from malignant nodules. Consequently, a percutaneous biopsy is often necessary to determine if a nodule is benign or malignant. We hypothesized that deep learning in conjunction with professional image characterization could improve nodule characterization and reduce benign biopsies. We extracted our features using convolutional auto-encoders, local binary patterns as well as histogram of oriented gradients descriptors in association with medical professional thyroid image characterization. The experiment showed the classifiers using these features can improve negative predictive value of thyroid nodule evaluation using ultrasound.