The Value of Deep Learning in Differentiating Thyroid Adenomatoid Nodules on Ultrasound: A Dual-Center Study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Ultrasound, Fuzhou University Affiliated Provincial Hospital, Fuzhou 350000, China (S.C., X.L., X.T.Z., N.L., G.S.D., L.Z., S.Q.C.); Department of Ultrasound, Fujian Provincial Hospital, Fuzhou 350000, China (S.C., X.L., X.T.Z., N.L., G.S.D., L.Z., S.Q.C.); Department of Ultrasound, Shengli Clinical Medical College of Fujian Medical University, Fuzhou 350000, China (S.C., X.L., X.T.Z., N.L., G.S.D., L.Z., S.Q.C.).
- Department of Ultrasound, First Affiliated Hospital of Fujian Medical University, Fuzhou 350005, China (Z.L.Y.).
- Department of Ultrasound, Fuzhou University Affiliated Provincial Hospital, Fuzhou 350000, China (S.C., X.L., X.T.Z., N.L., G.S.D., L.Z., S.Q.C.); Department of Ultrasound, Fujian Provincial Hospital, Fuzhou 350000, China (S.C., X.L., X.T.Z., N.L., G.S.D., L.Z., S.Q.C.); Department of Ultrasound, Shengli Clinical Medical College of Fujian Medical University, Fuzhou 350000, China (S.C., X.L., X.T.Z., N.L., G.S.D., L.Z., S.Q.C.). Electronic address: [email protected].
Abstract
Follicular neoplasms are difficult to classify by ultrasound, as they often present as thyroid adenomatoid nodules (TANU). Furthermore, fine-needle aspiration cytology (FNAC) exhibits limited accuracy in differentiating the nature of follicular lesions. This study aimed to develop and validate a novel model to optimize the workflow for distinguishing benign from malignant TANU. This study enrolled 648 TANUs, which were divided into the training (Train), validation (Val), external test (Test), and surgically confirmed (SC) cohorts. Handcrafted radiomics features were extracted using PyRadiomics, and deep features were obtained using pretrained ResNet models. Robust features were selected via LASSO regression. The deep learning radiomics (DLR) signature integrated the deep transfer learning (DTL) and radiomics signatures. The combined signature integrated the DLR and clinical signatures. The performance of different signatures was evaluated using the Area Under the Curve (AUC). We also compared the diagnostic performance of our best signature with FNAC. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations were used to interpret the deep learning component. The DTL signature achieved the highest AUC of 0.959 in the Test cohort (the DLR, combined, radiomics, and clinical signatures were 0.937, 0.930, 0.835, and 0.599, respectively). Furthermore, the DTL signature outperformed FNAC in diagnostic accuracy, with higher specificity (96.2%) and overall accuracy (89.3%). Grad-CAM confirmed that key image regions significantly contributed to the signature's decisions. By eliminating the need to integrate clinical and radiomics features, the DTL signature streamlines the diagnostic process while maintaining robust performance, making it a highly effective tool for clinical decision-making.