Thyroid Ultrasound Diagnostic Report Generation Method Using Large Language Models

Thyroid ultrasound is an essential modality for disease detection; however, its diagnostic efficacy is frequently compromised by inconsistencies in image interpretation and reporting. To address this, this study introduces a multimodal framework that leverages large language models (LLMs) to improve the efficiency, accuracy, and standardization of thyroid ultrasound report generation. The proposed framework integrates a multi-path visual encoder for hierarchical feature extraction and employs a two-stage pipeline that utilizes MiniCPM-V 2.6 and Deepseek-R1 70B to generate structured reports. The fine-tuning of MiniCPM-V 2.6 was conducted using Low-Rank Adaptation (LoRA), which substantially reduced training costs while enhancing performance. Evaluated on a dataset of 8,312 TI-RADS-compliant thyroid ultrasound samples, the model demonstrated a 25.7% improvement in feature extraction accuracy. Furthermore, it achieved 92.8% accuracy in key terminology generation and 87.8% precision for identifying pathological elements. Most notably, the TI-RADS classification accuracy increased significantly from 22.1% to 82.3%. A comparative analysis confirms that Deepseek-R1 70B surpasses other leading models across all evaluation metrics. This framework represents a significant advancement in intelligent ultrasonic diagnosis and holds considerable potential to improve clinical outcomes.