Deep learning-based computer-aided diagnosis for parotid gland tumors on MRI.
Authors
Affiliations (7)
Affiliations (7)
- Department of Otolaryngology-Head and Neck Surgery, Ehime University Graduate School of Medicine, Shitsukawa, Toon, Ehime, 791-0295, Japan; Department of Head and Neck Surgery, Shizuoka Cancer Center, Shimonagakubo, Nagaizumi, Sunto District, Shizuoka, Japan.
- Department of Otolaryngology-Head and Neck Surgery, Ehime University Graduate School of Medicine, Shitsukawa, Toon, Ehime, 791-0295, Japan. Electronic address: [email protected].
- Graduate School of Science and Engineering, Ehime University, Do-go Himata, Matsuyama, Ehime, Japan.
- Department of Radiology, Ehime University Graduate School of Medicine, Shitsukawa, Toon, Ehime, Japan.
- Department of Head and Neck Surgery, National Cancer Center Hospital East, Kashiwanoha, Kashiwa, Chiba, Japan.
- Department of Otolaryngology-Head and Neck Surgery, Ehime University Graduate School of Medicine, Shitsukawa, Toon, Ehime, 791-0295, Japan.
- Department of Head and Neck Surgery, Shizuoka Cancer Center, Shimonagakubo, Nagaizumi, Sunto District, Shizuoka, Japan.
Abstract
Accurate imaging-based differentiation of benign and malignant parotid tumors is essential for determining appropriate therapy. We investigated whether a deep learning (DL)-based computer-aided diagnosis (CAD) system provides incremental clinical value by improving reader performance in MRI-based evaluation of parotid gland tumors. This study included patients who underwent surgery between 2000 and 2022. A DL model (EfficientNet-based CNN) was developed using MRI data from 170 histologically confirmed cases; model evaluation used five-fold cross-validation with patient-wise splits to avoid leakage. In each fold, training, validation, and test subsets were exclusive, and the model output was the probability of malignancy. For the reader study, we pooled the test subsets from all five folds, yielding 134 cases for analysis. Four readers classified tumors on MRI (benign vs malignant) first without and then with CAD. We evaluated diagnostic accuracy, sensitivity, specificity, and AUC with and without CAD. The DL model achieved an accuracy of 0.85 and an AUC of 0.93. With CAD, readers showed better performance than without CAD (accuracy: 0.86 vs 0.76; AUC: 0.94 vs 0.82; p < 0.001). This improvement was observed in both experienced radiologists (AUC: 0.85 to 0.94) and residents (AUC: 0.79 to 0.95). In subgroup analyses, CAD significantly improved the predicted malignancy probability for intermediate- and high-grade tumors and pT2-T4 tumors, but not for low-grade or pT1 tumors. DL-based CAD improved diagnostic performance regardless of reader experience and may be particularly valuable for high-grade and locally advanced parotid gland tumors in MRI-based assessment.