Knowledge-driven deep learning: automated ultrasound diagnosis of follicular thyroid carcinoma

Background: Follicular thyroid carcinoma accounts for 10-15% of thyroid cancer, and its misdiagnosis rate of fine needle aspirations is as high as 50%. At present, there are few reports on the automatic classification of benign and malignant follicular nodules. It is of great significance to obtain objective diagnostic opinions through ultrasound deep learning to assist physicians to make accurate diagnosis of follicular thyroid cancer. Methods: Different from the conventional methods, SRI-LDF (segmented ROI image fused with LBP and DWT), which employs channel stacking to fuse spatial, frequency and texture features, is developed to achieve data enhancement in this paper. Specially, knowledge driven deep learning framework (KDDL), which includes the knowledge guide module (KGM) and follicular nodule capsule networks (FNCN) is created to classify the benign and malignant thyroid nodules. The accuracy of benign and malignant classification of thyroid and follicular nodules was verified on two of our own datasets: TND-PUH3 (3771 images, 1304 benign, 2467 malignant), which includes a subset of 607 follicular nodule images (458 benign, 149 malignant), and TNVD-PUH3 (882 images, 447 benign, 253 malignant) specifically for follicular nodules. Results: The proposed KDDL achieved an AUROC of 93.7% on the whole nodule dataset, and for the follicular nodules sets TNVD-PUH3, AUROC score was 90.7%, which performed better than the existing CNN-based method. Conclusions: Our proposed nodules differential diagnosis method presents superiority and is promising in providing clinical assistance in difficult tasks such as follicular nodule diagnosis.