A Multi-Task Intelligent Grading System for Thyroid Ultrasound Images

To solve the clinical problem of intelligent grading of nodules in thyroid ultrasound images under the C-TIRADS standard, a multi-task intelligent grading method was proposed to improve feature recognition ability and address data quality issues. An improved encoding-segmentation-classification multi-task framework was used to extract conventional features such as texture and shape of different diagnostic indicators, which were fused with features extracted based on deep learning. An adaptive loss function and noise rejection algorithm were designed to solve the problems of sample imbalance and incorrect labeling. On a clinical dataset containing 3857 samples, the classification accuracy of diagnostic indicators and the grading accuracy of nodules were 84.39% (an increase of 2.81% compared with the baseline DenseNet121) and 80.74% (an increase of 3.46%), respectively. The experimental results show that this method can accurately diagnose clinical indicators such as orientation, composition, echogenicity, margin characteristics, and calcification. It is reliable for auxiliary diagnosis of thyroid nodules and has high clinical application value.