Thyroid nodule risk assessment based on neural network text classification method

Abstract Thyroid nodules represent a prevalent endocrine disorder, and neck ultrasonography stands as a widely utilized and efficacious diagnostic modality within clinical practice. Sonographers ascertain the malignant tumor risk level by integrating various attributes of thyroid nodule features, including shape, capsule characteristics, aspect ratio, and calcification patterns.To mitigate the potential inadequacies and biases associated with manual assessments, this study introduces a text classification method grounded in deep neural networks to facilitate risk assessment for malignant thyroid nodules based on ultrasound findings. Three distinct classification models, TextCNN, bidirectional Long Short-Term Memory (LSTM), and CNN combined with Gated Recurrent Unit (GRU), were trained and evaluated using actual ultrasound data samples. Remarkably, the CNN + GRU model exhibited an evaluation accuracy exceeding 95%, underscoring the feasibility and effectiveness of employing neural network-based text classification for thyroid nodule risk assessment through segmental ultrasonography.Furthermore, the experiment's findings underscore that this approach possesses remarkable clinical applicability without reliance on word segmentation techniques or specialized dictionaries.