Classification of Glioma Grades on Diffusion-Weighted MRI Using Random Forest in Machine Learning

Background: Glioma, a type of brain tumor, is classified into two grades, LGG and HGG, and requires early detection due to the high mortality rate from brain tumors in Indonesia, with 5,323 new cases and 4,229 deaths. For patients with kidney disorders, gadolinium-based contrast MRI is unsuitable, making DWI images and AI-based systems viable alternatives for detection. Objective: This study aims to develop a machine learning (ML) model using the random forest (RF) algorithm for automatic glioma grade classification from DWI MRI Brain images and to evaluate its performance. Method: Purposive sampling was used to collect 1,848 DWI MRI brain images. A quasi-experimental design with a post-test control group was employed. The model's validity was assessed by media experts using a Likert scale to evaluate aspects such as user interface design, system performance, flexibility, usability, and reliability. Reliability was tested using Fleiss' Kappa for inter-rater reliability. Results: The ML model achieved 82% accuracy and a micro-AUC of 0.96, excelling in Normal and Grade 4 classifications but needing better recall for Grades 1 and 2. IT experts rated it positively: 84% for User Interface Design, 82% for Usability, 84.4% for System Performance, and 68.9% and 73.3% for Flexibility and Reliability. The Wilcoxon test found no significant differences from respondents (p > 0.05), with Fleiss' Kappa at 0.85 and 92% Observed Agreement. Conclusion: This study successfully developed and tested an RF model for glioma classification, demonstrating consistent and accurate results.