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

Reflectance confocal microscopy (RCM) enables noninvasive, cellular-resolution optical biopsies of human skin by acquiring \emph{en-face} images at successive depths. However, RCM stacks remain highly anisotropic, as the lateral resolution is approximately 0.5 μm, while axial resolution is limited by optical sectioning, making cross-sectional interpretation difficult. In this work, we aim to computationally densify axial sampling and estimate the diffraction-limited RCM section corresponding to an unobserved intermediate focal depth. We introduce CD-RCM, an RCM-specific framework for continuous-depth novel-view synthesis from sparse confocal stacks. CD-RCM formulates axial stack densification as a feedforward, depth-queryable synthesis problem, predicting intermediate optical sections at arbitrary unseen depths without per-stack optimization. Unlike conventional neural rendering methods designed for surface-level multi-view imagery with geometric parallax, CD-RCM exploits the structured axial acquisition geometry of RCM, where slices share a fixed lateral field of view and differ primarily by focal depth. We approximate this process as a virtual pinhole camera undergoing pure axial translation, enabling Plücker ray embeddings to condition a decoder-only transformer on observed and queried depths. This formulation allows CD-RCM to learn cross-stack priors over depth-dependent tissue morphology while treating axial position as a continuous, queryable dimension. To better preserve RCM-specific micro-texture, we further introduce a skin-specific perceptual feature loss using a domain-adapted self-supervised ViT trained on RCM mosaics. Experiments on 216 \emph{in vivo} RCM stacks show that CD-RCM outperforms B-spline, cubic-spline, Gaussian, and bilinear interpolation across photometric and perceptual metrics, achieving 23.36 PSNR, 0.581 SSIM, and 0.314 LPIPS. We further compare against learned alternatives, mainly a NeRF-like [1] per-stack coordinate MLP that substantially underperforms classical interpolation, while FILM [2], evaluated both zero-shot and after RCM fine-tuning, fails to reliably preserve depth-dependent morphology. These results indicate that the gains do not arise merely from applying a learned model, but from the RCM-specific continuous-depth formulation. Qualitatively, CD-RCM better preserves fine cellular morphology and produces smoother sagittal, coronal, and oblique virtual sections. Tenfold densification of a full stack requires approximately 0.8 seconds on a single A6000 GPU. Overall, CD-RCM establishes continuous-depth reconstruction in optical sectioning microscopy as a promising neural view-synthesis problem and motivates further investigation of amortized, physics-informed models for computational microscopy. References [1] Mildenhall, Ben, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. "Nerf: Representing scenes as neural radiance fields for view synthesis." Communications of the ACM 65, no. 1 (2021): 99-106. [2] Reda, Fitsum, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. "Film: Frame interpolation for large motion." In European Conference on Computer Vision, pp. 250-266. Cham: Springer Nature Switzerland, 2022.