Success of machine learning methods in brain tumour detection

Objective: The aim of this study is to evaluate the performance of machine learning models in detecting brain tumors by testing them on MRI images of pediatric patients who were diagnosed with tumors by radiology specialists. Materials and Methods: This methodological study retrospectively evaluated brain MRI images of pediatric patients under the age of 18 who applied to the tertiary healthcare facility of Dicle University Faculty of Medicine Hospital between 2010 and 2024. A total of 460 patients were included, consisting of 230 patients diagnosed with brain tumors and 230 without. In total, 1,380 MRI images were analyzed. Patients were divided into two groups: tumor-positive and tumor-negative. Deep learning-based architectures, YOLOv12 and EfficientNetB4, were utilized for tumor localization and binary classification in MRI images. The Python programming language was used for image analysis, and the Roboflow platform was employed for image labeling. Results: In the classification of tumor and non-tumor MRI images using the CNN model, EfficientNetB4 achieved an accuracy rate of 97%. The YOLOv12 model, on the other hand, reached an mAP50 accuracy of 89%. Throughout the training process, both models demonstrated stable loss reduction and consistent validation performance. The model produced reliable results on both the training and test datasets, and no overfitting was observed. Conclusion: The findings of this study suggest that the developed models can serve as supportive tools in clinical decision-making processes, assisting radiologists while offering advantages in terms of workload reduction and cost efficiency.