Multi-Dataset Evaluation of Hybrid Models for Brain Tumor Diagnosis

Abstract

Brain tumor detection is the identification and categorization of aberrant brain tissues using methods like MRI for tumor diagnosis and tracking. This sophisticated technique uses deep learning to analyze images, resulting in precise early detection and treatment. This study uses hybrid architectures for different deep-learning applications to offer a comprehensive hybridization strategy with promising prospects for improving the diagnostic precision of images obtained for medical diagnostics. In this study, it employs three separate datasets, previously known as Brain MRI images, Br35H and BraTS, to assess several architectures, including ResNet, VGG, Inception, EfficientNet, DenseNet121, MobileNetV2, Xception, NASNetMobile, and InceptionResNetV2. For the Brain MRI dataset, the findings demonstrated that the VGG16 model had a training accuracy of 99.93% with the lowest train loss of 0.0238; on all three datasets, the InceptionV3 showed exceptional robustness, with an accuracy of 99.78%. Although hybrid models that combined architectures such as Xception, NASNetMobile, and InceptionResNetV2 performed effectively, they also appeared to overfit, with validation and test losses being comparatively larger than training accuracy. The hybrid model hybrid model (EfficientNet, DenseNet121, MobileNetV2) achieved 99.87% training accuracy using the BraTS dataset. These findings indicate the possibility of applying deep learning architectures more effectively to better diagnose brain tumors, in addition to rigorous model optimization and selection to reduce the tendency for overfitting. This study encourages hybrid DL models' application in the medical area.