A Two-Stage Hybrid U-Net and Optimized SE-MobileNet Architecture for High- Precision Neuro-Ancillary Segmentation

Abstract

Background: High-fidelity delineation of multi-parametric pathomorphological zones and fine neuro-ancillary structures within structural magnetic resonance imaging (MRI) is a critical bottleneck for defining targeted stereotactic radiation field trajectories and planning interventions in neuro-oncology. Conventional single-stage deep architectures frequently encounter feature extraction breakdowns when processing low-contrast tissue interfaces and severe intensity variations across imaging modalities.
Methods: This study introduces a mathematically unified, two-stage hybrid machine learning pipeline tailored for automated, high-precision neuro-ancillary tissue segmentation using real-time dataset cohorts. The first stage deploys an enhanced multi-scale convolutional U-Net designed to isolate the coarse region-of-interest masks while stripping away non-parenchymal artifacts. The second stage applies an optimized Squeeze-and-Excitation Mobile-Net (SE-Mobile-Net) infrastructure that introduces localized channel-attention blocks to selectively recalibrate intermediate feature vectors and track fine-grained structural boundary gradients.
Results: Validation protocols executed on real-time multi-modal brain tissue datasets demonstrate that the dual-stage framework achieves state-of-the-art performance. The model reached a mean Dice Similarity Coefficient (DSC) of 95.14% for intricate target tissue classes, combined with a significant spatial reduction in the 95th percentile Hausdorff Distance (HD95) to 1.98 mm.
Conclusion: Empirical evaluations verify that pairing localized channel attention with sequential region-of-interest filtering effectively suppresses voxel-level classification errors without expanding computational latency. The system sets a highly reliable and computationally efficient baseline ready for seamless deployment into computer-aided clinical neuro-diagnostics.