Comparative Analysis of Transfer Learning Models for Brain Tumor Detection

Brain tumors rank among the most dangerous and life-threatening conditions, requiring accurate and timely detection for efficient treatment. Conventional diagnosis techniques, which rely on radiologists’ manual interpretation of medical pictures, might be time-consuming and susceptible to inconsistencies. To address these issues, deep learning (DL) models have become effective instruments for automating tumor classification, improving both efficiency and accuracy. In this manuscript, we evaluate the efficiency of various DL models in classifying MRI images as either tumorous or non-tumorous using the BR35H dataset, which includes 2,100 training images and 600 validation images. We assess nine pre-trained transfer learning (TL) models: DenseNet121, MobileNetV2, Inception-V3, MobileNetV3Large, Xception, DenseNet201, VGG16, AlexNet, InceptionResNetV2, and VGG16. Each model is fine-tuned and evaluated using precision, accuracy, recall, and F1-score metrics. Among all the techniques, Xception achieved the greatest accuracy at 98.33%, while MobileNetV3Large exhibited the lowest accuracy at 75.16%. This outcome showed the efficiency of advanced DL models in binary tumor classification, highlighting the importance of model selection in developing accurate and reliable diagnostic tools. Overall, the Xception model stands out as the most promising for automating the detection of tumorous and non-tumorous conditions in MRI images, offering a significant improvement over the traditional diagnostic method.