Two Eyes Are Better Than One: How a Neural Network Learns to See Tumors express
A doctor studies a brain scan: somewhere in the tangle of tissue, a tumor is hiding. One glance picks out small spots and edges, another takes in the overall shape and location. A person combines both, but across hundreds of scans that gets tedious. A new program does the same thing automatically.
One part of it works like a magnifying glass: it slides over the image and gathers fine details — texture, edges, local changes. The second part takes a wide view, as if from above, and connects distant regions: where the tumor sits relative to the ventricles, how it fits into the overall picture. Previously, such views were combined crudely, just mixing the results. Now a 'negotiation room' has been added: a special module decides how much to take from each view for a given scan and a given region.
The model was trained on a set of scans with and without tumors. It correctly identified tumors in 97.6% of cases — higher than either part alone and higher than previous fusion methods. Precision, recall, and the overall quality metric all improved. This means that flexibly weighting the two views helps see the disease more clearly.
This approach could be useful not only for brain tumors: any medical images where both fine details and the big picture matter could benefit from 'two eyes with a conductor.'
Original abstract
as written by the authors arXiv:2604.23137
Early detection and classifying brain tumors using Magnetic Resonance Imaging (MRI) images is highly important but difficult to extract in medical images. Convolutional Neural Networks (CNNs) are good at capturing both local texture and spatial information whereas Vision Transformers (ViTs) are good at capturing long-range global dependencies. We propose a new hybrid architecture that combines a SqueezeNet-style CNN branch with a MobileViT-style global transformer branch, through an Adaptive Attention Gate mechanism, in this paper. The gate learns dynamically per-sample, per-feature weights to weight the contribution of each branch, allowing context-sensitive merging of local and global representations. The proposed model has a test accuracy of 97.60, a precision of 97.30, a recall of 97.50, an F1-score of 97.40, and a macro-average area under the curve (AUC) of 0.9946 with a trained and evaluated on the Brain Tumor MRI Dataset (Kaggle). These scores are higher than single CNN and ViT baselines, and current competitive fusion methods, showing that dynamic feature weighting is an effective way to classify medical images.
An AUC of 0.9946 means that if you pick a random scan with a tumor and a random one without, the model will almost always give the first one a higher 'tumor-likeness' score.