Deep Brain Net: An Optimized Deep Learning Model for Brain tumor Detection in MRI Images Using EfficientNetB0 and ResNet50 with Transfer Learning

Early and accurate brain tumour classification from MRI is challenging due to inter-scanner variability, ambiguous boundaries, and limited labelled data. This paper presents DeepBrainNet, a lightweight yet powerful hybrid model that combines EfficientNetB0 and ResNet50 with transfer learning and fuzzy C-means guided feature selection. Three public MRI sources—figshare, Br35H, and SARTAJ—were harmonized through de-duplication and label audits, with patient-wise stratified splits to ensure reliability. A brief hyperparameter search identified Adam optimizer with learning rate 1e-4, batch size 32, cosine learning rate decay, label smoothing, and early stopping as the best settings. On the combined dataset, DeepBrainNet achieved an accuracy of 88.9 percent, a weighted F1-score of 89%, and a macro-AUC of 98%. The model demonstrates strong performance and scalability while maintaining computational efficiency. The study also outlines a Big Data pathway through mixed-precision and distributed training, reporting throughput, latency, and memory utilization as key system metrics.