Multimodal Convolutional Neural Network Models Allow for the Accurate Classification and Grading of Preoperative Meningioma Brain Tumors

Magnetic resonance imaging (MRI) and computed tomography (CT) scans are vital for diagnosing brain tumors, but human error, image subtleties, cyst growth, and nuances in World Health Organization (WHO) grading can impede accuracy. Invasive biopsies remain the only definitive method for meningioma diagnosis. Convolutional Neural Networks (CNNs), machine learning models used in image classification, offer a promising solution. By fine-tuning the pre-trained CNN EfficientNetB0 on various preoperative brain tumors and meningioma subtypes, image-based diagnosis can become more robust and accurate. In this study, two CNN models either classified or graded multimodal CT and MRI images. One dataset included tumor types (meningioma, glioma, pituitary, cysts, or none), while the other had images WHO graded one to three. The data, from accurately annotated and diverse open-source databases, was normalized, augmented, and stripped of excess information. Additionally, class-average and Focal Tversky Loss were included to assess and reduce incorrect outputs. Results were analyzed using accuracy, f1, recall, precision, loss, confusion matrices, Receiver Operating Characteristic (ROC) analysis, and attention studies. Both CNNs achieved over 98% accuracy with high recall and precision scores. ROC area under the curve (AUC) scores above 0.978 indicated strong class discrimination. The attention study indicated focus on tumor mass instead of extraneous variables. Multimodal CNNs, particularly the EfficientNetB0 model, are potential alternatives to invasive biopsies and human evaluation. Their capability to handle complex meningioma cases suggests promising avenues for other tumor types or diagnostic modalities at a cheap cost.