Development and internal validation of a multimodal MRI-FNAC radiomics-pathomics model for predicting cervical lymph node metastasis in papillary thyroid carcinoma.

Abstract

Background: Cervical lymph node metastasis (CLNM) occurs in 40-60% of papillary thyroid carcinoma (PTC) cases and critically affects surgical planning and prognosis. Current preoperative assessment relies mainly on ultrasound and ultrasound-guided fine-needle aspiration cytology (FNAC), which have limited sensitivity and a high false-negative rate. These shortcomings frequently lead to either missed metastases or unnecessary prophylactic neck dissection with risks of recurrent laryngeal nerve injury and hypoparathyroidism. Therefore, more accurate non-invasive preoperative prediction tools are urgently needed. This study aimed to develop and internally validate a multimodal magnetic resonance imaging (MRI)-FNAC radiomics-pathomics model based on deep learning features for predicting CLNM in patients with PTC.

Methods: We retrospectively enrolled 128 PTC patients who underwent thyroidectomy and neck dissection (78 with CLNM, 50 without). Patients were randomly split into training (n=90) and test (n=38) sets at a 7:3 ratio. Deep features were extracted from preoperative MRI using a pretrained ResNet-50 model and from FNAC hematoxylin and eosin (H&E) images. Six machine learning algorithms were applied to develop models based on MRI radiomics, pathomics, and multimodal fusion features. Performance was evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), decision curve analysis (DCA), DeLong test, and calibration metrics (Brier score and Hosmer-Lemeshow test).

Results: The cohort had a mean age of 48.82±10.02 years (71.1% female). No significant differences were found between CLNM-positive and -negative groups in age, sex, or tumor diameter, whereas multifocality and extrathyroidal extension were more prevalent in the CLNM-positive group. In the test set, the multimodal k-nearest neighbors (KNN) model achieved the highest AUC of 0.872 (95% confidence interval: 0.778-0.955). DeLong tests confirmed its superiority over the best pathomics-only model (P=0.046) and the best MRI-only model (P=0.003). The multimodal model also demonstrated the best calibration (Brier score 0.1059; Hosmer-Lemeshow P=0.72) and superior clinical net benefit on DCA.

Conclusions: The multimodal MRI-FNAC radiomics-pathomics model showed promising performance for predicting CLNM in PTC. However, external validation in larger multicenter cohorts is required before clinical implementation.