PriorPath

Synthetic histology is the standard answer to annotation scarcity in computational pathology, and the usual recipe runs in two stages: sample a binary tissue mask from noise, then translate it into a photorealistic stained image. The weakness is the first stage. A generator asked to invent tissue layouts from noise settles into one region of the space of real ones — a t-SNE projection of its masks forms a tight cluster inside the far broader cloud of masks from real biopsies — so the synthetic slides reproduce the narrowness they were meant to cure.

PriorPath, from the Network Biology Research Lab at Technion working with the Division of Pathology at Rambam Health Care Campus, replaces the noise with a prior the user draws. A pathologist sketches a coarse binary mask — roughly where tissue sits and where background sits — and a conditional GAN translates it into a fine-grained semantic mask with realistic tissue geometry. A second conditional GAN renders that mask as a stained image patch. Because the sketch is an input rather than a sample, the spatial arrangement of the result is specified rather than drawn, and the generated masks spread across the real mask space instead of concentrating in one corner of it.

The design descends from DEPAS, the same group's earlier mask generator, which injected spatial noise into hidden layers but still sampled from noise. PriorPath keeps the mask-then-image structure and changes only what conditions the first stage. The work was accepted to IEEE EMBC 2025.

#Key Features

• Drawn priors instead of sampled ones: the coarse mask is a hand sketch made on a tablet in ordinary painting software, so a user specifies the tissue distribution they want.
• Two-stage cascade: pix2pix maps coarse mask to fine mask and pix2pixHD maps fine mask to a 512×1024 RGB patch, yielding both the semantic label map and the image.
• Covers the semantic mask space: masks from varied sketches populate the real mask distribution rather than collapsing into one mode, measured by t-SNE coverage and by FID within each cell of a 3×3 grid over that space.
• Two staining modalities: trained separately on H&E and on PD-L1 immunohistochemistry, so the approach is not tied to one stain chemistry.

#Technical Details

The coarse-to-fine stage is pix2pix — a U-Net generator with skip connections and a PatchGAN discriminator — trained on paired coarse and fine masks. Fine masks come from thresholding grayscale patches to separate tissue from air; their coarse partners are produced by a morphological open-then-close pass with 5×5 and 10×10 kernels. CycleGAN was trained as an unpaired alternative and dropped: pix2pix beat it on FID by a factor of 1.316 on H&E and 1.378 on IHC. The rendering stage is pix2pixHD with a residual-convolution generator and two multiscale discriminators. Training ran on a single RTX A6000 with Adam, batch size 1, and a learning rate of 2e-4 held constant for 50 epochs then decayed to zero over 50 more.

Four models were trained, one per cancer realization: TCGA prostate adenocarcinoma, skin cutaneous melanoma and lung squamous cell carcinoma in H&E, plus an in-house non-small cell lung carcinoma PD-L1 IHC cohort from Rambam. Each drew on 50 whole slide images (55 for IHC) cut into 512×1024 patches — roughly 5,559–5,983 training patches per H&E cohort and 4,000 for IHC, with 1,000 held-out test patches (1,500 for IHC). PriorPath reached FID 71.246 on PRAD, 61.892 on SKCM, 60.782 on LUSC and 103.348 on NSCLC, against 198.704, 138.604, 437.812 and 138.629 for DEPAS, with matching gains in Kolmogorov–Smirnov distance and KL divergence. The coarse-mask corpus is about 100 pathologist sketches per cancer type, augmented into a larger repository.

#Applications

The immediate use is augmenting training sets for downstream pathology models — tissue segmentation, tumor detection — in organs or stains where annotated whole slide images are limited or skew toward particular presentations. Because the label map is generated alongside the image, each synthetic patch arrives with its ground truth attached, which is the expensive part of a real annotation. Pathologist control of the layout also allows a set to be enriched deliberately for arrangements the real cohort underrepresents.

#Impact

PriorPath reframes de novo mask generation as image-to-image translation rather than sampling — a small architectural change with a direct effect on mode collapse, the failure that limited its predecessor. Its components are off-the-shelf; the contribution is the conditioning formulation and the hand-drawn coarse-mask corpus. Its reach is bounded in three ways the authors state plainly: the masks are binary tissue-versus-background rather than the multi-label cellular maps that would let a user place specific cell types, each cancer type and stain needs its own trained model with no demonstrated transfer between organs, and evaluation uses held-out patches from the same cohorts rather than an external one. No code, weights, or checkpoints have been released; data is offered on request from the corresponding author.