A Hybrid Convolutional Neural Network and Extreme Value Theory Model for Meningioma Filtering in Open Set Brain Tumor Data

Meningioma image filtering frequently encounters difficulties when presented with data that deviates from the training distribution. Challenges may emerge when addressing novel tumor types or non-tumor images. This paper presents a hybrid strategy that integrates Convolutional Neural Network (CNN) and Extreme Value Theory (EVT) models to filter out-ofdistribution (OOD) data. This study classifies meningioma as the in-distribution (ID) category. Three CNN models, specifically EfficientNetV2B0, MobileNetV2, and InceptionV3, were employed as feature extractors to produce confidence ratings, which were later examined using EVT at the distribution's tail. The test results demonstrate that InceptionV3 is very effective, reaching a validation accuracy of 99.5% and an F1-score of 83% in closed-set situations and 88% in open-set situations. MobileNetV2 showed outstanding efficiency, getting an F1-score of 97% on the testing meningioma classification and a recall of 99% for detecting OOD cases. Models displaying strong generalization ability demonstrated the effectiveness of EVT integration in improving OOD performance. This study shows that pairing a reliable CNN model with the right EVT setup can make a meningioma filtering system more trustworthy in unpredictable situations.