Hard Spatial Gating for Precision-Driven Brain Metastasis Segmentation: Addressing the Over-Segmentation Paradox in Deep Attention Networks

TL;DR

Brain metastasis segmentation is hindered by extreme class imbalance and very small lesions, which cause soft-attention models to exhibit high recall but poor precision and boundary accuracy. The authors introduce SG-Net, a hard spatial gating architecture that uses grouped feature processing and binary-like gating to aggressively suppress background while preserving tumor features, achieving a Dice of 0.5578 and a 56.13 mm HD95 with significantly fewer parameters (0.67M). This precision-first approach outperforms Attention U-Net and ResU-Net in key metrics and demonstrates substantial improvements in boundary delineation, crucial for stereotactic radiosurgery planning. The work underscores the importance of task-specific architectural biases for small-lesion segmentation and highlights potential for efficient deployment in resource-limited clinical settings, with implications for broader applications in precision medicine.

Abstract

Brain metastasis segmentation in MRI remains a formidable challenge due to diminutive lesion sizes (5-15 mm) and extreme class imbalance (less than 2% tumor volume). While soft-attention CNNs are widely used, we identify a critical failure mode termed the "over-segmentation paradox," where models achieve high sensitivity (recall > 0.88) but suffer from catastrophic precision collapse (precision < 0.23) and boundary errors exceeding 150 mm. This imprecision poses significant risks for stereotactic radiosurgery planning. To address this, we introduce the Spatial Gating Network (SG-Net), a precision-first architecture employing hard spatial gating mechanisms. Unlike traditional soft attention, SG-Net enforces strict feature selection to aggressively suppress background artifacts while preserving tumor features. Validated on the Brain-Mets-Lung-MRI dataset (n=92), SG-Net achieves a Dice Similarity Coefficient of 0.5578 +/- 0.0243 (95% CI: 0.45-0.67), statistically outperforming Attention U-Net (p < 0.001) and ResU-Net (p < 0.001). Most critically, SG-Net demonstrates a threefold improvement in boundary precision, achieving a 95% Hausdorff Distance of 56.13 mm compared to 157.52 mm for Attention U-Net, while maintaining robust recall (0.79) and superior precision (0.52 vs. 0.20). Furthermore, SG-Net requires only 0.67M parameters (8.8x fewer than Attention U-Net), facilitating deployment in resource-constrained environments. These findings establish hard spatial gating as a robust solution for precision-driven lesion detection, directly enhancing radiosurgery accuracy.

Figures (7)

• Figure 1: The proposed workflow. Multi-modal MRI inputs are preprocessed and patched before being fed into SG-Net. The model optimizes a compound loss function to generate precise binary segmentation masks.
• Figure 2: Visual assessment of segmentation quality across two different patients. Rows represent distinct test cases. Columns (L-R): T1-Weighted MRI, FLAIR sequence, Ground Truth mask, and Overlay of the tumor region. SG-Net effectively localizes lesions in both cases.
• Figure 3: Architecture of SG-Net. The model uses a 3D U-Net backbone with Spatial Gating Modules (SGM) in skip connections to filter noise. The SGM (right) uses spatial attention to refine features.
• Figure 4: Distribution of Dice Similarity Coefficients across test subjects. The white diamonds ($\diamondsuit$) indicate mean values, while dots represent outliers. SG-Net demonstrates a consistently higher mean and median Dice score compared to baseline models.
• Figure 5: The trade-off between Sensitivity and Precision. (a) While Attention U-Net achieves high recall, (b) it suffers from extreme boundary errors (High HD95). SG-Net achieves the lowest HD95, indicating precise localization.