Research map / 3D vision / NeRF and radiance fields

CD-RCM: Generalizable Continuous-Depth Novel View Synthesis for Reflectance Confocal Microscopy

NeurIPS 2026 · 3D vision › NeRF and radiance fields · approach: Inverse rendering with neural fields · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CD-RCM, a feedforward decoder-only transformer that synthesizes unseen intermediate depth slices from sparse reflectance confocal microscopy z-stacks, with a DINOv3-based skin-specific perceptual loss.

Problem

Reflectance confocal microscopy stacks are anisotropic and sparsely sampled in depth, limiting interpretation of skin; conventional interpolation lacks anatomical priors and per-stack optimization of neural rendering is costly.

Approach

Model RCM axial acquisition as a virtual pinhole camera translating along z, tokenize input slices and Plücker target ray embeddings into a LVSM-style decoder-only transformer that renders arbitrary intermediate depths in one forward pass, supervised by a LoRA-adapted DINOv3 skin-specific perceptual loss.

Key results

• CD-RCM achieves PSNR 23.36, SSIM 0.581, LPIPS 0.314, L_SPF 0.145, outperforming B-spline (22.02 PSNR, LPIPS 0.375), cubic spline, and Gaussian interpolation.
• 10x densification of a full stack takes 0.8 seconds on a single A6000 GPU.
• In the ablation, the single-layer skin-specific perceptual loss with LPIPS gives the best results (PSNR 23.35, LPIPS 0.288) versus VGG perceptual loss plus LPIPS (22.65, 0.329).

Surprising findings

• Gaussian interpolation gets higher PSNR and SSIM than some learned variants but worse perceptual scores, which the authors attribute to blur favoring PSNR.
• A multi-layer skin-specific perceptual loss underperforms the single-layer final-layer design across all metrics.

Limitations

• Residual misalignment after SIFT-based registration degrades reconstruction in some stacks.
• Dataset is small and limited to healthy skin from a single imaging device, with no pathological cases.
• Synthesized slices are not directly observed and risk hallucinated structures being misread clinically.

Experimental setup

Datasets: in vivo RCM z-stacks (216 stacks, VivaScope 1500) RCM mosaic crops for skin-specific feature extractor

Benchmarks: in-house RCM stack test split (60 stacks)

Metrics: PSNR SSIM LPIPS L_SPF

Baselines: B-spline interpolation Cubic spline interpolation Gaussian interpolation

Builds on: LVSM DINOv3 LPIPS FlashAttention-2

Closest papers by meaning

• From Pixels to Views: Learning Angular-Aware and Physics-Consistent Representations for Light Field Microscopy (NeurIPS 2025 · similarity 0.85)
• S^3-Face: SSS-Compliant Facial Reflectance Estimation via Diffusion Priors (CVPR 2025 · similarity 0.82)
• SR-CACO-2: A Dataset for Confocal Fluorescence Microscopy Image Super-Resolution (NeurIPS 2024 · similarity 0.82)
• V2V3D: View-to-View Denoised 3D Reconstruction for Light Field Microscopy (CVPR 2025 · similarity 0.82)
• Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.82)
• MicroDiffusion: Implicit Representation-Guided Diffusion for 3D Reconstruction from Limited 2D Microscopy Projections (CVPR 2024 · similarity 0.82)