Using Machine Learning on MRI Radiomics to Diagnose Parotid Tumours Before Comparing Performance with Radiologists: A...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Parotid gland tumors, both benign and malignant, pose diagnostic challenges.
• Current methods like MRI and biopsy have limited accuracy, often necessitating surgery for diagnosis.

Purpose of the Study:

• To develop a machine learning algorithm using MRI characteristics for automated parotid tumor classification.
• To compare the algorithm's performance against junior and senior radiologists to assess its clinical utility.

Main Methods:

• Utilized radiomics features extracted from four MRI sequences of 134 patients with parotid tumors.
• Trained random forest and logistic regression models to predict histopathological subtypes.
• Conducted a clinical experiment comparing algorithm-assisted diagnosis with radiologist diagnoses.

Main Results:

• The random forest model achieved 0.720 accuracy, 0.860 specificity, and 0.720 sensitivity for all subtypes (AUC 0.838).
• For benign vs. malignant discrimination, the algorithm reached 0.760 accuracy and 0.769 AUC.
• Junior radiologists' diagnostic sensitivity and accuracy improved by 6% when using the proposed model.

Conclusions:

• Machine learning with radiomics shows promise in differentiating parotid tumors, potentially reducing diagnostic surgery.
• The algorithm can aid in physician training and improve diagnostic accuracy, especially for less experienced radiologists.