Classification of parotid gland tumors by using multimodal MRI and deep learning

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Parotid gland tumors require accurate differentiation for effective treatment.
• Magnetic Resonance Imaging (MRI) sequences like T2-weighted, postcontrast T1-weighted, and diffusion-weighted images show potential for tumor discrimination.
• Accurate classification of parotid tumors is crucial for patient management.

Purpose of the Study:

• To develop and evaluate a fully automatic deep learning system for diagnosing parotid gland tumors using multimodal MRI data.
• To compare the classification accuracy of different MRI contrast combinations for parotid tumor diagnosis.
• To assess the performance of a deep learning model in differentiating Warthin tumors, pleomorphic adenomas, and malignant tumors.

Main Methods:

• A U-Net convolutional neural network was employed for segmentation and classification of parotid gland tumors.
• The U-Net model utilized transfer learning and optimized batch distribution for improved accuracy.
• Five combinations of MRI contrasts were used as input data to evaluate classification performance.

Main Results:

• The deep learning model incorporating diffusion-related MRI parameters outperformed models using only structural MR images.
• The diffusion-based model achieved an accuracy of 0.81 for Warthin tumors, 0.76 for pleomorphic adenomas, and 0.71 for malignant tumors (n=85).
• Sensitivity and specificity varied across tumor types, with the model successfully classifying Warthin tumors and pleomorphic adenomas but not malignant tumors.

Conclusions:

• The proposed deep learning model demonstrates effectiveness in classifying benign parotid gland tumors (Warthin tumor, pleomorphic adenoma).
• The model's performance in differentiating malignant parotid tumors requires further improvement.
• Diffusion-weighted MRI parameters are valuable for deep learning-based parotid tumor classification.