研究目的
To establish an automatic image recognition and diagnosis system for thyroid nodules in ultrasound imaging using the YOLOv2 neural network, evaluate its performance compared to radiologists, and investigate its clinical application value.
研究成果
The AI system shows comparable sensitivity and accuracy to radiologists for malignant nodules and higher specificity for benign nodules, aiding in reducing false positives and unnecessary procedures. It offers real-time detection, potentially saving diagnosis time and reducing subjective errors, with significant clinical application potential, though further multicenter studies and integration are needed.
研究不足
1. Majority of malignant nodules were papillary thyroid carcinoma (240/242), lacking diversity in malignancy types. 2. Did not evaluate enhancement of AI system on radiologists' diagnoses. 3. Single-center study with potential selection bias. 4. System not yet integrated into ultrasound equipment for full clinical assessment.
1：Experimental Design and Method Selection:
An end-to-end detection network based on YOLOv2 integrated with Resnet v2-50 was used for simultaneous identification of nodule location and type without manual labeling. K-means clustering determined prior predictive frames, and transfer learning was employed with pretraining on VOC dataset and fine-tuning on thyroid ultrasound images.
2：Sample Selection and Data Sources:
2450 benign and 2557 malignant thyroid nodule images from patients of different ages and genders were collected and labeled for training. For testing, 351 nodule images and 213 normal images from 276 patients (53 males, 223 females, average age 46.3 years) were retrospectively selected from the Affiliated Hospital of Qingdao University, with pathological diagnosis as gold standard.
3：3 years) were retrospectively selected from the Affiliated Hospital of Qingdao University, with pathological diagnosis as gold standard.
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: Ultrasound machines: GE Logiq E8 (GE Healthcare), Philips iE Elite, Philips iU22 (Philips Healthcare) with high-frequency linear array probes (frequencies 6-15 MHz, 3-11 MHz, 5-12 MHz). Software: SPSS version 19.0, MedCalc for Windows version 15.0 for statistical analysis.
4：0, MedCalc for Windows version 0 for statistical analysis.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Ultrasound examinations performed by senior radiologists; images input into AI system for automatic recognition and diagnosis. Radiologists diagnosed based on TI-RADS; AI system output bounding boxes and classification. Nodule localization assessed visually as excellent, satisfactory, or poor.
5：Data Analysis Methods:
Chi-square test for differences in diagnoses; sensitivity, specificity, PPV, NPV, accuracy calculated; ROC curves and AUC compared using DeLong method with p<0.05 significance.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
