Brain tumor classification using optimized ResNet50 with dynamic precision optimization for enhanced speed and diagnostic accuracy.

The detection and classification of brain tumors remain a major challenge in the medical field due to their morphological complexity. In this study, by optimizing a model from the ResNet family and applying architectural modifications, including the addition of specialized layers, an efficient system was designed that simultaneously reduced computational parameters, decreased training time, and increased inference speed, while improving diagnostic accuracy up to 99.69%. The proposed model, through the use of Transfer Learning (TL) and Fine-Tuning techniques, as well as the implementation of intelligent computational precision management, is capable of maintaining a balance between model complexity and operational accuracy, enabling deployment on resource-limited hardware. The model achieved 100% precision in detecting glioma, pituitary, and healthy cases, and 98.71% precision in identifying meningioma. To enhance the interpretability of results, the model's output was combined with a Random Forest (RF) classifier, allowing physicians to analyze the decision-making process and the key features influencing diagnosis. The experiments were conducted on the brain MRI dataset, comprising 7023 images categorized into four classes. The dataset was split into 80% training and 20% testing, and a 5-fold cross-validation strategy was also applied to ensure model generalization. The obtained results demonstrate the high efficiency, computational optimization, and clinical applicability of the proposed architecture in developing intelligent automated brain tumor diagnosis systems.