Cloud-Based Artificial Intelligence Classification of Common Intracranial Tumors on Magnetic Resonance Imaging

Brain tumors comprise a heterogeneous group of neurological conditions for which timely and accurate classification is important for diagnostic evaluation, treatment planning, and prognosis. Magnetic resonance imaging is the principal imaging modality used to evaluate intracranial tumors; however, image interpretation may be time-consuming and dependent on specialist expertise. This study aimed to develop and internally evaluate a cloud-based artificial intelligence model for classifying glioma, meningioma, pituitary tumor, and no-tumor two-dimensional brain MRI images from a single publicly available dataset using image-level validation. A total of 5,335 images were obtained from a publicly available dataset and divided into training, validation, and testing subsets using an approximately 80-10-10 distribution. The dataset contained 1,400 glioma images, 1,292 meningioma images, 1,362 pituitary tumor images, and 1,281 no-tumor images. Model development was performed using the Google Cloud AutoML image-classification platform in a single-label multiclass configuration. Performance was evaluated using overall and class-specific average precision, precision, recall, precision-recall curves, confidence-threshold analysis, and a multiclass confusion matrix. At a confidence threshold of 0.50, the model achieved an overall average precision of 0.986, precision of 99.2%, and recall of 99.1%. Class-specific average precision was 1.000 for glioma, 1.000 for no-tumor images, 0.986 for pituitary tumors, and 0.973 for meningiomas. Correct classification rates ranged from approximately 99% to 100% across the four categories. These findings demonstrate the feasibility of using an accessible cloud-based AutoML platform for brain MRI image classification. These findings demonstrate strong internal image-level classification performance; however, the inability to confirm patient-level independence, absence of external validation, and use of a single public dataset limit interpretation of the reported metrics. Independent patient-level and multi-institutional validation is required before the model's diagnostic performance or clinical applicability can be established.