Anatomically-Guided Attention for False Positive Reduction in Brain Tumor Segmentation

Abstract U-Net-based brain tumor segmentation can produce false positives in anatomically implausible locations, which inflate tumor volume estimates and reduce clinical trust. We investigate whether injecting an anatomical location prior into the network's attention can reduce these errors. We propose a prior informed attention gate that conditions U-Net features on empirical tumor-location prior maps computed from the training annotations, applied as a residual spatial gate at the bottleneck and at every decoder stage. Using the BRISC dataset of 2D contrast-enhanced T1-weighted MRI slices (glioma, meningioma and pituitary tumor), we compare the proposed model with a vanilla U-Net and a learned-gate ablation under identical training, using 3-fold cross-validation and a held-out test set of 860 images. We evaluate Dice, IoU, precision, recall, Hausdorff distance, HD95 and a dedicated far false-positive measure. The proposed model matched the baseline in Dice (0.709 versus 0.708) and showed a slightly lower but non-significant HD95 (36.5 versus 37.9 px, p = 0.26); it did not significantly reduce far false positives (49.7 versus 48.0 px, p = 0.82). A simple location prior injected through a spatial gate is therefore insufficient on BRISC, and stronger prior integration, longer training and additional datasets are needed to establish any benefit.