Diagnostic Performance of Machine Learning Models Based on ADC Maps in Parotid Gland Tumors

Abstract

Objective: This study aims to evaluate the diagnostic performance of radiomics features extracted from ADC maps for differentiating benign from malignant parotid tumors using machine learning models, demonstrating their potential as supplementary diagnostic tools. Materials and Methods: Forty-six patients who underwent diffusion-weighted imaging (DWI) between January 2018 and December 2022 and had histopathologically confirmed parotid tumors were retrospectively analyzed. Patients were classified as pleomorphic adenoma (n=15), Warthin tumor (n=19), and malignant tumor (n=12). Lesions were manually segmented on ADC maps to construct a radiomics dataset, and various regression methods and ML algorithms were applied for both benign–malignant differentiation and subtype classification. Results: Stable radiomics features selected by the Fast Correlation-Based Filter (FCBF) method were used to assess the performance of multiple ML algorithms. The Artificial Neural Network (ANN) achieved the highest accuracy in distinguishing benign from malignant tumors, demonstrating the potential of ML models like Neural Networks and SVM to support diagnostic decisions. Conclusion: The ANN algorithm showed promising performance in differentiating parotid tumor groups. ADC-based radiomics and machine learning may provide noninvasive diagnostic support; however, these methods should be considered adjuncts to, rather than replacements for, histopathological diagnosis.