Bilateral disease in the classic subtype of papillary thyroid carcinoma: clinical significance and development of an artificial intelligence-based multimodal prediction model.

BACKGROUND

Papillary thyroid carcinoma (PTC) is the most common thyroid malignancy, with multifocality being a frequent occurrence. The clinical significance of bilateral disease versus unilateral multifocality remains controversial. This study aimed to evaluate the clinicopathological characteristics and recurrence patterns of bilateral disease in patients with the classic subtype of PTC, and to develop a multimodal artificial intelligence model for preoperative identification of bilateral disease in this specific histological subtype.

METHODS

Clinical data and ultrasound images of 1,218 patients with PTC were retrospectively collected from two hospitals. Clinicopathological features and recurrence patterns were analyzed. Four prediction models were developed: a logistic regression model with clinical parameters (Model A), a machine learning-based model with LASSO-selected features (Model B), a deep learning framework using TresNet architecture (Model C), and an integrated model combining Models B and C (Model D). Model performance was assessed in internal and external validation cohorts, with clinical utility evaluated across different radiologist experience levels.

RESULTS

Among the 1,098 patients from the primary institution, 26.1% had multifocal disease, of which 62.7% had bilateral involvement. Bilateral disease was associated with larger tumor size, higher rates of lymph node metastasis, and worse recurrence-free survival compared to unifocal disease and unilateral multifocality. Multivariate analysis identified bilateral disease as an independent risk factor for recurrence (HR = 9.664, P = 0.005). The integrated Model D achieved superior performance with AUC of 0.970 in the training set, 0.932 in the validation set, and 0.848 in the external validation set. In the external validation group of 120 patients, Model D's implementation significantly improved diagnostic accuracy across all radiologist experience levels, with junior radiologists benefiting most substantially (+20.9% improvement).

CONCLUSION

In classic subtype PTC, bilateral disease, rather than multifocality alone, is a significant risk factor for recurrence. The developed multimodal artificial intelligence model effectively predicts bilateral disease preoperatively, potentially guiding surgical planning and improving diagnostic accuracy irrespective of radiologist experience.