Bezier-YOLO enables real-time brain tumor segmentation in multi-plane MRI through boundary-aware parametric contour regression

Accurate brain-tumor segmentation requires boundary predictions that follow curved, irregular lesion margins while remaining fast enough for practical use. We present Bezier-YOLO, a real-time detection-and-segmentation model that represents each detected tumor as a closed cubic-Bézier contour rather than a box-centered mask. The standard YOLO regression branch is replaced by a parametric head predicting 24 control points, and a compact Channel–Spatial Attention Module refines backbone features with limited overhead. Training combines control-point regression, soft polygon-IoU supervision, and tangent-continuity regularization. Experiments used the tumor-positive segmentation subset of BRISC2025, a public collection of contrast-enhanced T1-weighted brain MRI slices comprising glioma, meningioma, and pituitary tumor cases with non-empty masks; no-tumor images were excluded because they do not provide contours for fitting. Bezier-YOLO-l achieved 95.3% mask [email protected], 70.7% mask [email protected]:0.95, and 84.0% Dice at 72 FPS on an RTX 3090, outperforming the evaluated YOLO segmentation baselines under the same protocol. Because every model was trained and tested under one protocol on a single public dataset, these differences are reported descriptively rather than as statistically established superiority. The largest gains occurred in mask-level agreement, whereas failures were concentrated in irregular or multifocal gliomas. The study is deliberately scoped to a slice-level, tumor-positive setting, so specificity on normal anatomy, boundary-distance agreement, and cross-center robustness were not assessed. These results indicate that compact contour regression is a promising route for real-time MRI tumor segmentation, although external validation and normal-case false-positive analysis remain necessary.