Preliminary machine-learning model with clinical, US, and CEUS features for the diagnosis of thyroid follicular-patterned lesions.
Authors
Affiliations (4)
Affiliations (4)
- Department of Ultrasound in Medicine, Shanghai Sixth People's Hospital, Shanghai Institute of Ultrasound in Medicine, 600 Yishan Road, Shanghai, 200233, China.
- School of data science, Fudan University, 220 Handan Road, Shanghai, China.
- Department of Ultrasound in Medicine, Shanghai Sixth People's Hospital, Shanghai Institute of Ultrasound in Medicine, 600 Yishan Road, Shanghai, 200233, China. [email protected].
- Department of Ultrasound in Medicine, Shanghai Sixth People's Hospital, Shanghai Institute of Ultrasound in Medicine, 600 Yishan Road, Shanghai, 200233, China. [email protected].
Abstract
Though follicular thyroid carcinoma can be confirmed postoperatively by the histological findings of capsular or vascular invasion, preoperative diagnosis of follicular-patterned lesions has long been a diagnostic challenge. This study seeks to establish a machine-learning (ML) model based on clinical, US, and CEUS features for the differential diagnosis of thyroid follicular-patterned lesions (TFPLs). Patients with surgical pathologically confirmed TFPLs who underwent preoperative US and CEUS between January 2013 to April 2023 were enrolled in this retrospective study. We utilized five ML algorithms (logistic regression, random forest [RF], k-nearest neighbor [KNN], support vector machine [SVM], and elastic net [EN]) to construct an optimized model via US, CEUS and clinical data for the differential diagnosis of TFPLs. Model performance was evaluated with sensitivity, area under the precision-recall curve (AUPRC), and area under the receiver operating characteristic curve (AUROC). 114 patients were finally included. The sensitivities of the five ML algorithms (Logistic, RF, KNN, SVM, EN) were 0.93, 0.97, 0.93, 0.97, 0.93 for the training set, and 0.40, 1.00, 0.80, 1.00, 0.60 for the test set. The AUPRCs were 0.91, 0.97, 0.99, 0.91, 0.91 for the training set, and 0.65, 0.71, 0.38, 0.57, 0.65 for the test set. The AUROCs were 0.95, 0.98, 0.99, 0.89, 0.93 for the training set. When applied to the test set, the RF model had a significantly higher AUROC value (0.92; 95% CI: 0.88, 0.96) than other ML algorithms (0.91, 0.89, 0.91, 0.91, P < .05) with significant features including peripheral halo sign, thyroglobulin, rim enhancement, peak intensity and wash-out time. Our ML model integrating clinical, US, and CEUS features achieved high sensitivity for preoperative differentiation of TFPLs, potentially guiding surgical planning-a step toward clinical use that requires CEUS standardization and external validation. This study was registered at www.chictr.org.cn (no. ChiCTR2200066254, date: November 29, 2022).