Tan Dianhuan, Zhai Yue, Hu Zhengming, Xu Bingxuan, Zheng Tingting, Chen Yan, Sun Desheng
Shenzhen Key Laboratory for Drug Addiction and Medication Safety, Department of Ultrasound, Institute of Ultrasonic Medicine, Peking University Shenzhen Hospital, Shenzhen Peking University-The Hong Kong University of Science and Technology Medical Center, Shenzhen, China.
Quant Imaging Med Surg. 2025 Jun 6;15(6):5719-5738. doi: 10.21037/qims-24-2336. Epub 2025 May 30.
Accurately differentiating thyroid conditions, particularly malignant and benign nodules, is crucial for effective treatment planning. Fine needle aspiration biopsy (FNAB), the standard diagnostic method for suspected malignancies detected by ultrasound, is invasive and can be unnecessary. This study introduces a novel dual-stage deep learning architecture for four-class classification of thyroid nodules, aiming to improve diagnostic accuracy and reduce unnecessary procedures. The framework integrates segmentation and classification stages, offering more precise and consistent analyses than traditional binary classification, potentially improving guidance for FNAB decisions and reducing invasiveness.
Our dual-stage deep learning framework integrates segmentation and classification. The segmentation stage uses K-Net to delineate nodule boundaries. The classification stage employs MobileViT to classify nodules into four categories: papillary carcinoma, medullary carcinoma, nodular goiter with adenomatous hyperplasia, and chronic lymphocytic thyroiditis (CLT).
K-Net achieved a mean intersection over union of 87.06% in segmenting the nodule boundaries. MobileViT achieved a mean accuracy of 92.33% on the validation set and 90.27% on the test set. The confusion matrix demonstrated high diagonal values, indicating accurate classification across all categories. The model demonstrated a precision of 91.30% and recall of 89.76% on the validation set. Performance was particularly strong for papillary thyroid carcinoma and CLT.
This approach offers an efficient tool for analyzing thyroid nodules, potentially improving guidance for FNAB decisions and reducing invasiveness. The framework shows promise for clinical application, offering a non-invasive method to aid in the differential diagnosis of thyroid nodules. By integrating advanced segmentation and classification, the model offers a refined approach to thyroid nodule analysis.
准确区分甲状腺疾病，尤其是恶性和良性结节，对于有效的治疗规划至关重要。细针穿刺活检（FNAB）是超声检测出疑似恶性肿瘤的标准诊断方法，具有侵入性且可能不必要。本研究引入了一种用于甲状腺结节四类分类的新型双阶段深度学习架构，旨在提高诊断准确性并减少不必要的程序。该框架整合了分割和分类阶段，比传统的二元分类提供更精确和一致的分析，可能改善对FNAB决策的指导并降低侵入性。
我们的双阶段深度学习框架整合了分割和分类。分割阶段使用K-Net描绘结节边界。分类阶段采用MobileViT将结节分为四类：乳头状癌、髓样癌、伴有腺瘤样增生的结节性甲状腺肿和慢性淋巴细胞性甲状腺炎（CLT）。
K-Net在分割结节边界时平均交并比达到87.06%。MobileViT在验证集上的平均准确率为92.33%，在测试集上为90.27%。混淆矩阵显示对角线值较高，表明所有类别分类准确。该模型在验证集上的精确率为91.30%，召回率为89.76%。对甲状腺乳头状癌和CLT的表现尤为出色。
这种方法为分析甲状腺结节提供了一种有效的工具，可能改善对FNAB决策的指导并降低侵入性。该框架显示出临床应用的前景，提供了一种非侵入性方法来辅助甲状腺结节的鉴别诊断。通过整合先进的分割和分类，该模型为甲状腺结节分析提供了一种精细的方法。
