A Hybrid Approach for Brain-Tumor Detection and Classification from MRI Images

Many existing deep-learning models for brain-tumor diagnosis report very high accuracy, yet they often suffer from overfitting, limited generalizability, and poor robustness, which reduces their clinical reliability. First, a large and diverse dataset was assembled from multiple publicly available sources to reduce data bias. Next, a series of regularization and calibration techniques were applied, including label smoothing, mixup augmentation (α=0.4), exponential moving average (EMA) of model weights to prevent overfitting, stabilize training, and temperature scaling to calibrate predicted probabilities. Since brain-tumor diagnosis directly affects critical clinical decisions, both model accuracy and reliability of its predictions are essential. For that, six diverse, powerful pretrained backbones (ConvNeXt, EfficientNetV2, DenseNet201, Swin, ViT, and DINOv3) were used as frozen feature extractors under pure transfer learning. On top of these frozen features, lightweight classification heads were trained using concatenation, attention, and gated fusion. Then, the predicted-class probabilities from the selected heads were combined and passed to a logistic-regression meta-learner to produce the final prediction, with the aim of improving accuracy and reliability. This architecture is computationally efficient because only lightweight classifiers and a logistic meta-learner are trained, while the heavy backbones remain frozen. The best-performing variant on TEST2, attn_fp_ens , was further evaluated using synchronized 10-fold cross-validation on the TRAIN set. The best-performing TEST variant, the gated-fusion ensemble with fingerprint features (gated_fp_ens), achieved 99.33% test accuracy and 99.28% macro-F1. Under domain shift on the external hospital dataset (TEST2), the attention-fusion ensemble with fingerprint features (attn_fp_ens) showed the strongest performance among the evaluated variants, achieving 95.16% accuracy and 94.11% macro-F1. The TEST2 dataset was collected from a local hospital for external evaluation.