Back to all papers

Ultrasound deep learning for predicting aggressiveness in papillary thyroid carcinoma: an exploratory analysis of transcriptomic correlates.

November 14, 2026pubmed logopapers

Authors

Jiang H,Lu W,Liu Z,Xie X,Xu X,Zhang C

Affiliations (3)

  • Department of Ultrasound, The First Affiliated Hospital of Anhui Medical University, Hefei 230022, Anhui, China.
  • Department of Ultrasound, The Second Hospital of Anhui Medical University, Hefei 230601, Anhui, China.
  • Department of Ultrasound, The First Affiliated Hospital of Anhui Medical University, Hefei 230022, Anhui, China. Electronic address: [email protected].

Abstract

Preoperative assessment of aggressiveness in papillary thyroid carcinoma (PTC) is important for individualized treatment decision-making. This study aimed to develop a deep learning (DL) model to predict PTC aggressiveness and, in an exploratory analysis, to assess associations between model-derived imaging phenotypes and transcriptomic pathway signatures. Preoperative ultrasound images from 1023 patients with PTC were used to develop a DL model for aggressiveness prediction, and clinical features were further integrated to construct a combined DL-Clinic model. Grad-CAM was applied to generate class-specific heatmaps, and mean Grad-CAM activation intensities were extracted from the whole lesion ROI and its core, marginal, and microcalcification regions. Patients were further stratified according to the DL score and regional activation intensity, followed by exploratory GSEA to assess transcriptomic pathway associations. The DL-Clinic model improved the AUCs to 0.834, 0.759, and 0.750 in the internal validation, external test, and prospective test sets, respectively. Multivariable analysis identified the DL score as the strongest independent predictor of PTC aggressiveness (OR = 3.498, p < 0.001). Grad-CAM analysis showed higher activation intensity in the ROI and all subregions of aggressive cases. Exploratory GSEA suggested that high-risk and high-activation phenotypes tended to be associated with enrichment of pathways related to tumor proliferation, extracellular matrix remodeling, and antigen processing, whereas low-risk and low-activation phenotypes were predominantly enriched in immune-related pathways. The DL-Clinic model demonstrated discriminatory performance for the preoperative prediction of PTC aggressiveness. Exploratory imaging-transcriptomic analyses suggested associations between model-derived imaging phenotypes and transcriptomic pathway signatures. Further prospective multicenter studies are needed to establish the model's clinical utility, and these exploratory imaging-transcriptomic associations require independent validation.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.