Thyroid cancer pathological classification in ultrasound images based on faster R-CNN network

Abstract

Objective: This study explored the performance of deep learning Faster R-CNN network for the automatic detection of nodules in thyroid ultrasound images and its efficacy in predicting different pathological classifications of thyroid malignant nodules. Methods:We retrospectively collected 1762 thyroid cancer ultrasound images from 548 patients, of which 80% of the enhanced ultrasound images were used for training and testing, and the remaining 20% were used for validation. The pathological classification of all nodules was confirmed by pathology. We also evaluated the effectiveness of Faster R-CNN models based on different backbone networks in predicting the pathological classification of thyroid cancer. Results:The Faster R-CNN model using ResNet50 as the backbone network in this study had an AUROC of 0.873 and an mAP of 84.04% for pathological classification of malignant thyroid nodules. The accuracy rate for Papillary Thyroid Carcinoma (PTC) was 84.89%, for Medullary Thyroid Carcinoma (MTC) was 89.41%, for Anaplastic Thyroid Carcinoma (ATC) was 82.38%, and for Follicular Thyroid Carcinoma (FTC) was 81.36%. The overall accuracy rate of the final classification was 84.59%. The recall rates of the best model were 87.37%, 85.80%, 83.48%, and 82.30% for PTC, MTC, ATC, and FTC, respectively. Conclusion:This study demonstrates that deep learning Faster R-CNN network can detect thyroid nodules and has good diagnostic efficacy in distinguishing pathological classifications of thyroid malignant nodules. It has great potential for clinical applications.