Abstract:
MRI plays a key role in screening and triaging of brain tumors, and models done on a specific site may fail when scanners / sequences or quality of acquisition is altered. We formulate a bare question, which is can smarter data augmentation make classifiers perform better, and test it. Our comparison of a progressively increasing number of MRI-aware augmentations to four- class slice classification (glioma, meningioma, pituitary, no-tumor) is based on three convolutional backbones of varying capacities (DenseNet-121 and MobileNetV3-Large with ImageNet pretraining) and three staged sets of these augmentations of varying strengths (mild, strong, advanced). We report the results of using a fixed 80/10/10 split, AdamW, class-balanced loss, and early stopping on a held-out test set with the results of accuracy, macro- precision/recall/F1 and per-class F1. The trends are evident: the augmentation alters the needle the most in the case of the smaller MobileNetV3-Large (macro-F1 0.772 to 0.897; accuracy 0.896) and the stronger DenseNet-121 to the same direction (macro-F1 0.941 to 0.946; accuracy 0.946). Excessive regularization, on the other side, undermines even the tiniest CustomCNN(macro-F1 0.868 dropping to 0.698724) and it is important to note that policy strength should be aligned with model capacity. In general, the optimal model is the DenseNet- 121 with strong/advanced augmentation, whose accuracy and macro-F1 are 0.946 and 0.914- 0.967, respectively.The implication is pragmatic: a thoughtfully selected augmentation recipe could increase CNN-based brain-tumor classifiers to be more robust to the inherent noise of real-world MRI particularly when dealing with small models.
