NeuroSeg: A Bi-Hemispheric Spectral-Attention U-Net Workspace with Signed Distance Boundary Supervision, Bayesian Uncertainty Calibration, and Cross-Modal Neuro-Hematology Fusion

In this work, we developed NeuroSeg, an integrated clinical neuro-oncology imaging, uncertainty-quantification, and multimodal blood-biomarker fusion workspace for lower-grade and anaplastic glioma assessment. The system is organized as a four-tier architecture: a React 19 interactive Picture Archiving and Communication System (PACS) and clinical triage client, a FastAPI clinical validation and report-synthesis service tier, an in-process PyTorch neural inference and explainability tier, and a multimodal dataset and model-checkpoint persistence tier spanning N = 12,000 parametric records across 300 patients alongside 120 paired axial T2-FLAIR TIFF scans. In the spatial imaging plane, NeuroSeg extends a four-stage residual encoder-decoder backbone (43,991,373 trainable parameters) with 2D Real Fast Fourier Transform (rfft2) radial Gaussian high-pass Spectral Attention Gates (sigma = 0.08), an eight-head Bi-Hemispheric Contralateral Cross-Attention bottleneck that mirrors feature tokens across the sagittal midline, four-scale Deep Supervision (w = [0.5333, 0.2667, 0.1333, 0.0667]), and a GPU-accelerated iterative morphological Signed Distance Map (SDM) boundary loss (K = 8 steps). At inference time, the engine executes spatial Monte Carlo (MC) dropout (T in [2, 16] stochastic forward passes) to decompose voxel-wise predictive Shannon entropy and epistemic mutual information into a visual boundary uncertainty halo, paired with higher-order Grad-CAM++ saliency maps on decoder stage dec1 and 3D RECIST 1.1 volumetric reconstruction. In the systemic hematology plane, a Gated Multimodal Unit (CrossModalGatedFusionNet) fuses an 8-dimensional MRI volumetric/boundary vector with a 14-dimensional hematology and astroglial liquid-biopsy vector—comprising Hemoglobin, WBC, ANC, ALC, Platelets, Prothrombin INR, D-Dimer, Serum Sodium, Fasting Glucose, hs-CRP, Serum LDH, Serum GFAP, Neutrophil-to-Lymphocyte Ratio (NLR), and Systemic Immune-Inflammation Index (SII)—while enforcing deterministic Common Terminology Criteria for Adverse Events (CTCAE v5.0) Temozolomide (TMZ) dosing gates and a five-point Pre-Craniotomy Neurosurgical Clearance shield. Across a five-stage architectural ablation study, NeuroSeg improved segmentation Dice Similarity Coefficient (DSC) from 0.8310 (vanilla U-Net baseline B0) to 0.9184 (+8.74%), Intersection over Union (IoU) from 0.7425 to 0.8512, 95th-percentile Hausdorff Distance (HD95) from 8.64 mm to 3.42 mm (-5.22 mm), and Expected Calibration Error (ECE) from 0.0680 to 0.0192 (-71.8%). On the held-out multimodal test split (N_test = 1,500), cross-modal gated fusion achieved 94.33% four-class differential diagnosis accuracy (disambiguating Diffuse Glioma, Pyogenic Cerebral Abscess, Primary CNS Lymphoma, and Tumefactive Demyelination vs. 76.40% for unimodal B0), 99.87% WHO CNS Grade stratification accuracy, 0.9640 macro ROC-AUC, and zero false-clearance safety leaks across N_adv = 40 adversarial hematologic and out-of-distribution upload injections.