Deep Learning-Assisted MRI for Differentiating Parotid Gland Tumors: Comparison with FNAB and Radiologic Assessment.
Authors
Affiliations (3)
Affiliations (3)
- Department of Otorhinolaryngology, Harran University Faculty of Medicine, Şanlıurfa 63290, Türkiye.
- Department of Computer Engineering, Faculty of Engineering and Architecture, Mardin Artuklu University, Mardin 47100, Türkiye.
- Department of Radiology, Mehmet Akif İnan Training and Research Hospital, Şanlıurfa 63300, Türkiye.
Abstract
<b>Background/Objectives:</b> To evaluate the diagnostic performance of an MRI-based deep learning (DL) model for differentiating benign and malignant parotid gland tumors and to compare its performance with radiologic assessment and fine-needle aspiration biopsy (FNAB), using histopathology as the reference standard. <b>Methods:</b> This retrospective single-center study included 144 consecutive patients with histopathologically confirmed parotid gland tumors who underwent parotidectomy between January 2020 and December 2024. Preoperative MRI examinations were analyzed using a ResNet50-based convolutional neural network incorporating a Convolutional Block Attention Module (CBAM). Model performance was evaluated using patient-level stratified 5-fold cross-validation and compared with MRI and FNAB using sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). <b>Results:</b> The DL model achieved a sensitivity of 79.3% (95% CI, 61.6-90.2), specificity of 93.9% (95% CI, 87.9-96.9), and an overall accuracy of 91.0% (95% CI, 85.3-94.5). MRI demonstrated a sensitivity of 86.2% and specificity of 93.0%, whereas FNAB achieved a sensitivity of 82.8% and specificity of 96.5%. The DL model showed consistent performance across the validation folds. <b>Conclusions:</b> MRI-based deep learning demonstrated high diagnostic performance for the preoperative classification of parotid gland tumors and may serve as a complementary decision-support tool alongside MRI and FNAB. Although promising, these findings are limited by the retrospective single-center design and the lack of external validation. Prospective multicenter studies are required before routine clinical implementation.