Selecting representative histologic sections for cost-efficient 3D spatial transcriptomics and tumor microenvironment reconstruction

Background: Three-dimensional spatial transcriptomics (ST) is essential for characterizing tumor-microenvironment (TME) heterogeneity but profiling every serial section is cost-prohibitive. Prior work has explored selecting subsets of sections for imputing ST at interleaved positions. However, these approaches do not identify which sections to profile to most efficiently preserve the 3D architecture. An optimal strategy must reinforce high-confidence inference in regions with similar composition while ensuring inclusion of morphologically distinct areas. We developed a computational framework to select histologic sections that optimally preserve 3D TME structure while minimizing the number requiring ST profiling. Methods: Four 3-mm FFPE colorectal cancer cores (two per patient) were collected to capture tumor-stroma interfaces and tertiary lymphoid structures. ~300 serial 5-µm H&E sections per core (~1500 µm depth) were imaged at 40x resolution; sections with sufficient tissue were retained (reported tissue loss at: TL 1150 µm, TR 350 µm, BL 1005 µm, BR 1100 µm). Retained sections were co-registered with VALIS. Eleven tissue classes-including tumor, desmoplastic stroma (DS), necrosis, smooth muscle, immune aggregates, and mucosa-were segmented using a graph neural network. For each section, we extracted Prov-GigaPath embeddings and tissue class area proportions. Pairwise similarities, incorporating histology features and z-distance, informed a facility location submodular optimization algorithm to select k representative sections (k=3-30). Selected subsets were used to reconstruct 3D TME structure via class-specific alpha-shape meshes and voxelization. Reconstruction fidelity was assessed using voxelized intersection-over-union (IoU) relative to full-section reconstructions. Results: Cores differed in composition-TL: normal mucosa 50%, smooth muscle 26%, stroma 11%, DS 10%; TR: tumor 49%, DS 49%; BL: smooth muscle 43%, tumor 32%, stroma 23%, inflammation 2%; BR: smooth muscle 71%, DS 14%, stroma 11%, tumor 3%. Weighted IoU increased with k; mixed-effects models showed each section improved IoU by 0.014 (p<2×10⁻¹⁶), with core-specific slopes TL 0.015, TR 0.019, BL 0.016, BR 0.013. Saturation occurred at TL 18, TR 12, BL 17, BR 16 sections, with maximum weighted IoUs of 0.49, 0.64, 0.68, and 0.50. Class-specific IoUs correlated with tissue class prevalence (ρ=0.81, p=3.6x10⁻⁵), with abundant classes reconstructing best. Conclusion: Section-selection algorithms can identify small subsets of H&E sections that preserve 3D TME structure for cost-efficient ST profiling. Although reconstruction fidelity reflects tissue class abundance, 15-20 sections sufficiently capture 3D architecture. Future work will refine selection strategies and integrate selected ST sections with interleaved H&E sections to enhance 3D reconstruction.