Brain tumors pose a significant health challenge, necessitating early and precise detection to improve patient outcomes. Traditional diagnostic methods, such as manual MRI analysis by radiologists, are time-consuming, subjective, and prone to variability, highlighting the need for automated solutions. This study proposes an AI-driven deep learning framework that integrates ResNet-50 for feature extraction, DenseNet-121 for classification, and U-Net for segmentation, forming a comprehensive pipeline for automated tumor detection. These models were chosen due to their superior feature representation, classification accuracy, and segmentation precision, making them well-suited for medical image analysis. To enhance model performance, an advanced preprocessing strategy incorporating skull stripping via HD-BET, normalization, and Nyul’s Histogram Matching-based augmentation was implemented. The impact of skull stripping was extensively evaluated on two benchmark datasets—FigShare and Kaggle—demonstrating a significant increase in classification accuracy from 94.85% to 98.25% on FigShare and from 95.42% to 98.56% on Kaggle. Additionally, computational efficiency was analyzed, demonstrating the feasibility of real-time clinical deployment by optimizing model execution times and reducing inference delays. The proposed approach also yielded substantial improvements in precision, recall, and AUC scores, underscoring the necessity of robust preprocessing in medical image analysis. These results establish the framework as a computationally efficient and explainable AI solution, enhancing diagnostic reliability and clinical applicability.
Buy article PDF