A Systematic Machine Learning Approach for Brain Tumor Classification Through Synthetic Control Data Generation to Prevent Bias and Ensure Fairness in Predictive Models Using MRI Imaging

A significant challenge in brain cancer diagnosis in Ecuador is the reliance on visual interpretation of magnetic resonance imaging (MRI) by specialists, a process that is time-consuming and prone to human error. Alternatively, biopsies are invasive and costly, limiting their accessibility. In this study, the CRISP-DM methodology is applied to develop a Deep Learning-based classification model to predict the presence of malignant tumors in MRI. The phases of the proposed method are: 1. Data Preparation Phase; the Brats 2023 Adult Glioma dataset is utilized for cancer patients, and synthetic samples are generated using data augmentation techniques to represent non-cancer patients, achieving a balanced dataset to promote fairness. It is worth highlighting that the models developed are custom-built and trained from scratch, without the use of transfer learning. 2. Classification Model Development Phase; three models are developed: Convolutional Neural Network (CNN), Residual Neural Network (ResNet), and Support Vector Machine (SVM). 3. Evaluation Phase; the models are evaluated using classical classification metrics: accuracy, precision, recall, and F1-score. The CNN achieved an accuracy of 99.77%, outperforming models such as SVM and ResNet. This work lays the groundwork for future research involving local datasets with images from Ecuadorian patients, both with and without cancer, to enhance the model’s generalization and applicability in real clinical settings, while also emphasizing the importance of fairness in diagnostic outcomes.