RNA-SpecFlow: RNA-Conditioned Spectral Flow for Histology Synthesis

RNA-SpecFlow is a continuous-latent FlowAR histology synthesis model that conditions adjacent-scale feature transitions on RNA, using five latent grids (1, 2, 4, 8, 16).

• Why it matters: Morphology-based evaluation of RNA-conditioned histology synthesis is sensitive to detector inputs and analysis choices, including how nuclear detection and segmentation agreement are measured.
• What they did: The model uses a Biological Latent Adapter plus a shared spectral fusion module that separates aligned features into low- and high-frequency bands, modulating each band with RNA and combining via residual fusion. It was analyzed on archived outputs from five TCGA cancer cohorts with 500 saved real/generated tile pairs.
• The result: On 12 public images with 303 annotated nuclei, twofold enlargement increases HoVer-Net predictions from 287 to 415 while mean panoptic quality changes from 0.673 to 0.458, and switching analysis pipelines reverses real/generated count differences in two cohorts.