PINN-FFD: physics- and frequency-informed one-stage detector for brain tumor detection in MRI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate brain tumor detection from magnetic resonance imaging (MRI) is essential for clinical treatment planning. Existing one-stage detectors are purely data-driven and may produce spatially fragmented or noise-sensitive predictions under limited training data. We propose PINN-FFD, a physics- and frequency-informed extension of YOLO that constructs a continuous tumor occupancy heatmap from bounding-box outputs and jointly optimizes (i) a Laplacian smoothness constraint inspired by physics-informed neural networks and (ii) a frequency-domain penalty that suppresses high-frequency spectral energy in the predicted tumor field. Unlike conventional PINN approaches that solve PDE-based reconstruction tasks, PINN-FFD imposes field-level priors directly on detection outputs without modifying the YOLO backbone. On a brain tumor MRI dataset of 500 annotated 2D slices (400/50/50 train/validation/test split), PINN-FFD achieves precision/recall/mAP@0.5 of 0.901/0.871/0.893, outperforming the YOLO11n baseline (0.797/0.878/0.853) by +13.0% in precision and +4.7% in mAP@0.5 while maintaining comparable recall (0.871 vs. 0.878) and reducing heatmap L2 error to 0.015. Ablation, hyperparameter, and computational analyses confirm the contribution of each module and the feasibility of lightweight clinical deployment.