Deep Learning Methods for Thyroid Imaging Segmentation: a Systematic Review

Abstract

Thyroid nodules are highly prevalent, and reliable imaging is essential for diagnosis and treatment. Deep learning (DL) has become the leading approach for automating thyroid ultrasound analysis. Objective: The purpose of this study is to present a systematic review of the state-of-the-art progress in applying deep learning methods to thyroid image segmentation and evaluate their clinical readiness. Method: We conducted a systematic review of 48 primary studies published between 2020 and 2025. workflow steps, DL families, datasets, and deployment contexts are analyzed. Result: Six DL families were identified: U-Net variants, Transformers, Boundary-aware, Weakly/Semi-supervised, Diffusion, and Multimodal fusion. These methods map to a four-step clinical workflow. While many achieved strong performance (Dice <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\approx 0.85-0.95$</tex>; AUC <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&gt;0.90$</tex>), most remain research-level (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 60 \%$</tex>), with fewer validated in hospitals (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 30 \%$</tex>) or pilots (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 10 \%$</tex>). Conclusion: DL has advanced thyroid segmentation considerably but faces challenges of limited datasets, reproducibility, and lack of prospective validation. Standardized benchmarks and multi-center studies are needed to move toward clinical adoption.

Ähnliche Arbeiten