Dual-Domain Spectral Vision Transformer with Asymmetric Boundary-Guided Diffusion for Ultra-High Resolution Panoptic Histopathology Segmentation

Precise panoptic and semantic segmentation of microscopic tissue structures, malignant glandular morphologies, and infiltrative tumor boundaries in gigapixel Whole Slide Images (WSIs) represents an indispensable prerequisite for computational oncology, digital pathology grading, and personalized therapeutic stratification. While deep convolutional neural networks (CNNs) and Vision Transformers (ViTs) have advanced automated histological analysis, current paradigms face fundamental technical bottlenecks: standard non-overlapping tokenization introduces severe spatial attenuation and high-frequency edge loss, windowed self-attention mechanisms fail to resolve multi-scale cellular pleomorphism across vast receptive fields, and deterministic decoders yield overconfident, blurry predictions along ambiguous stromal-epithelial transition margins. In this paper, we propose SpectralDiff-Trans, a novel dual-domain hierarchical vision transformer integrated with an asymmetric boundary-guided conditional diffusion refinement decoder for ultra-high-resolution panoptic histopathology segmentation. SpectralDiff-Trans introduces a 2D Learnable Wavelet-Fourier Spectral Transform (LWFST) encoder that decouples low-frequency macroscopic tissue architecture from multi-directional high-frequency cellular boundary gradients across spatial frequency domains. To capture multi-scale morphological context without quadratic complexity, we formulate a Linearized Criss-Cross Axial Attention (LCCA-Attention) bridge that propagates contextual dependencies across horizontal, vertical, and diagonal spatial trajectories. Crucially, rather than relying on deterministic softmax segmentation heads, we formulate boundary delineation as a conditional reverse-time stochastic differential equation (SDE) diffusion process, wherein a score-based boundary refinement network progressively denoises uncertain topological contours guided by edge energy fields. We conduct extensive experiments on three benchmark datasets: the GlaS Gland Segmentation Challenge, the multi-organ MoNuSeg nuclear segmentation dataset, and the PanNuke panoptic cancer histology benchmark (covering 19 distinct tissue types). SpectralDiff-Trans achieves state-of-the-art segmentation fidelity, reaching a mean Dice Similarity Coefficient (DSC) of 89.8%, a Panoptic Quality (PQ) of 68.4%, and an ultra-low 95% Hausdorff Distance (HD95) of 4.12 µm, outperforming competitive supervised ViTs (Swin-UNet, TransUNet, UNETR) and generative segmentation baselines (MedSegDiff) by +4.2% in DSC and reducing boundary localization errors by 41.8%. Comprehensive ablations, theoretical convergence analyses, and clinical pathologist Turing assessments validate that SpectralDiff-Trans provides robust, clinically actionable morphological delineation for next-generation digital pathology.