<![CDATA[Headline accuracy is not enough to move a deep-learning brain tumor magnetic resonance imaging (MRI) classifier into the clinic. A model also needs repeated-split validation, trustworthy probabilities, a statistical sanity check, and a small enough footprint for hospital hardware. This study cover all four in one framework. Five convolutional neural network (CNN) backbones (visual geometry group (VGG)-16, residual network (ResNet)-50V2, MobileNetV2, EfficientNetB0, and dense convolutional network (DenseNet)-121) are trained on 4,600 public brain MRI scans (2,513 tumor, 2,087 healthy) under a shared two-phase transfer-learning recipe. The usual single split is replaced by stratified 5-fold cross-validation (CV), with paired McNemar and DeLong tests and 1,000 bootstrap resamples. Calibration is judged by expected calibration error (ECE), Brier, reliability diagrams, and temperature scaling. VGG16 wins mean accuracy (98.72 ± 0.42%); ResNet50V2 has the tightest area under the curve (AUC) (0.9986 ± 0.0006); the two are statistically tied. MobileNetV2 is the best-calibrated model out of the box (ECE = 0.0021) with only 2.59 M parameters, making it the most deployable. A four-level confidence-based risk score turns calibrated probabilities into triage tags. This study also flag a measured 3.23% format-duplicate leakage in the public dataset and presents the framework as a radiologist co-pilot, not a prognostic model.]]>
Copyrights © 2026
