Answers are grounded in this paper and cite their sources.
Background: Magnetic resonance imaging (MRI) is the reference modality for detecting brain tumours, but manual reading needs specialist time that is scarce in many health systems. Public brain-MRI datasets used to develop automated classifiers often contain duplicate images, which can inflate reported accuracy. We assessed whether a small convolutional neural network, selected and evaluated with explicit statistical procedures on deduplicated data, can separate tumour from healthy brain MRI. Methods: We used the public Brain Tumor Classification (MRI) dataset (3,264 images). Perceptual hashing removed 867 duplicate copies and 39 images whose duplicates carried conflicting labels, leaving 2,358 unique images. These were split by stratified sampling into training (1650), validation (354) and test (354) sets. A statistical convolutional neural network (SCNN) was tuned by a full grid search over 108 configurations of learning rate, batch size and network depth, scored only on the validation set. The selected model was evaluated once on the test set, with Wilson, bootstrap and DeLong 95% confidence intervals (CIs), and by five-fold cross-validation. Results: 52 of 108 configurations, all with learning rates of 0.01 or 0.1, failed to learn. The selected SCNN (learning rate 0.001, batch size 20, four convolutional and two dense layers) classified 339 of 354 test images correctly: accuracy 95.76% (95% CI 93.13-97.42), sensitivity 95.82% (92.98-97.54), specificity 95.35% (84.54-98.72), Matthews correlation coefficient (MCC) 0.828 (0.745-0.908) and area under the ROC curve (AUC) 0.990 (0.982-0.998). Five-fold cross-validation gave lower and more variable estimates: accuracy 87.43% (SD 3.42), MCC 0.601 (SD 0.062) and AUC 0.930 (SD 0.030). Conclusions: A lightweight, statistically selected CNN detected brain tumours on deduplicated public MRI with high discrimination, but cross-validation showed that a single test split overstates performance. Interval estimates, deduplication and external validation should be standard in this literature.
No takes yet. Share an insight, caveat, or question.
Emmanuel et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: