Deep learning on contrast-enhanced computed tomography for parotid tumor classification: Providing crucial diagnostic information beyond clinical and radiologic evaluation.
TL;DR
Deep learning applied to CE-CT demonstrated strong diagnostic performance for the preoperative classification of PTs in this cohort and may serve as a powerful non-invasive adjunct to standard diagnostic modalities without adding procedural burden to the diagnostic workup.
Abstract
Purpose
Preoperative differentiation between benign and malignant parotid tumors (PTs) remains challenging despite clinical examination, cross-sectional imaging, and biopsy. Accurate malignancy assessment is crucial to optimize surgical planning and minimize morbidity.
Methods
We retrospectively analyzed 66 patients who underwent parotidectomy between 2008 and 2024, including 33 with malignant and 33 with benign PTs matched for demographics. All patients had contrast-enhanced computed tomography (CE-CT), and segmented tumor volumes were evaluated using a three-dimensional convolutional neural network. Diagnostic performance was compared with standard clinical and radiologic workup.
Results
Standard clinical and radiologic workup achieved a sensitivity of 60.6 %, increasing to 69.7 % with fine needle aspiration cytology (FNAC) or core needle biopsy (CNB). The deep learning model achieved an area under the ROC curve of 0.94, with both sensitivity and specificity exceeding 90 % at optimized thresholds, outperforming conventional diagnostics.
Conclusion
Deep learning applied to CE-CT demonstrated strong diagnostic performance for the preoperative classification of PTs in this cohort and may serve as a powerful non-invasive adjunct to standard diagnostic modalities without adding procedural burden to the diagnostic workup.