Résumé: The detection and brain tumor (BT) segmentation and classification are mandatory steps before any radiotherapy or surgery. When performed manually, segmentation is time-consuming and exposed to human errors. Therefore, significant efforts have been made to automate the process. In this study, a proposed automatic discriminative learning-based approach for brain tumor classification and segmentation using a metaheuristic optimizer called Sparrow Search Algorithm (SpaSA). The segmentation process is performed using UNet models (i.e., U-Net, U-Net++, Attention U-Net, and V-net). Additionally, the learning and SpaSA optimization is performed using pre-trained CNN models (i.e., MobileNet, MobileNetV2, MobileNetV3Small, MobileNetV3Large, EfficientNetB0, EfficientNetB1, EfficientNetB2, EfficientNetB3, EfficientNetB4, EfficientNetB5,VGG16, and VGG19). To optimize the training hyperparameters, the SpaSA metaheuristic optimizer is used. The dataset is collected from 6 public sources. Two types of datasets are generated. One with 2-classes and the other with 4-classes. The best-reported scores by U-Net architecture are 99.73% accuracy, 99.93% specificity, 99.35% AUC, 99.78% IoU, and 99.80% Dice for the whole tumor region. For the 2-classes dataset, the best reported overall accuracy from the applied CNN experiments is 99.99% by the MobileNetV3 Large pre-trained model. The average accuracy is 99.92%. Similarly, For the 4-classes dataset, the best reported overall accuracy from the applied CNN experiments is 99.73% by the EfficientNetB2 pre-trained model. The average accuracy is 99.19%. The suggested approach is compared with 11 related studies.

Langue: Anglais