DEPAS

Conditioning an image generator on a semantic mask is the cheapest way to get annotated histology: the synthetic patch arrives with its pixel-level label map already attached, which is the part of a real slide that costs a pathologist's time. The catch is where the masks come from. Tissue geometry — the irregular lacework of stroma and the air gaps between it — is not something a person sketches convincingly, so masks are harvested from real slides instead, which caps the number of synthetic images at the number of real annotations already in hand. Unconditional GANs scale without limit but emit pixels, not labels.

DEPAS breaks the cap by generating the masks themselves. A generator built on the DCGAN template decodes a latent Gaussian sample into a 512×1024 binary tissue-versus-air semantic mask; a separately trained pix2pixHD stage then renders that mask as a photorealistic stained patch. Once trained, the pipeline consumes nothing but noise, so mask-image pairs can be produced indefinitely. It comes from the Savir Lab in the Department of Physiology, Biophysics and Systems Biology at Technion with the Division of Pathology at Rambam Health Care Campus, and was published at IEEE EMBC 2023.

The technical obstacle is that a semantic mask is discrete while backpropagation is not: a step function at the generator's output severs the gradient path from the discriminator. DEPAS's answer is to anneal a continuous relaxation into a discrete one over training. The same group later replaced the noise input with a hand-drawn coarse mask in PriorPath, which trades unlimited sampling for user control over layout.

#Key Features

• Discrete Adaptive Block: the final layer is a sigmoid whose slope starts at 1 and is incremented every 10 epochs, sharpening toward a step function while staying differentiable throughout. The multi-label version uses a channel-wise softmax whose temperature is divided by 1.25 on the same schedule, mimicking an argmax.
• Spatial noise injection: a stock DCGAN draws a 1×1 latent vector per channel, which upsamples into visibly repetitive texture at high resolution. DEPAS also injects 2D Gaussian arrays into the decoder's hidden layers after a learnable scaling.
• Multi-scale discriminators: three discriminators see the mask at 100%, 50% and 25% resolution with equal loss weights, so tissue architecture and fine intercellular gaps are both judged.
• Multi-label semantic masks: beyond tissue/air, the generator can emit per-pixel PD-L1 attribute maps — PD-L1 positive, PD-L1 negative, inflammation, cells, air and other — letting a user request synthetic biopsies with a chosen cellular composition.

#Technical Details

The generator is five transpose-convolution blocks, each followed by batch normalization and ReLU, with scaled spatial noise added between them and the Discrete Adaptive Block on the end. Four independent models were trained, one per histology realization: TCGA prostate adenocarcinoma, skin cutaneous melanoma and lung squamous cell carcinoma in H&E at 50 whole slide images each, plus 27 in-house PD-L1 immunohistochemistry slides from non-small cell lung carcinoma patients at Rambam (IRB approval 0522-10-RMB). Slides were cut into 512×1024 patches, those with more than 85% background discarded, leaving 6,000 patches per H&E cohort and 2,012 for IHC at an 85/15 split. Binary ground-truth masks came from grayscale thresholding at 204 for H&E and 235 for IHC. Training used Adam at 2e-4 with betas 0.5/0.999, batch size 8 for 100 epochs, on a single RTX A6000.

Against a DCGAN baseline, DEPAS masks were closer to held-out real masks on Kolmogorov–Smirnov distance, KL divergence and Fréchet inception distance across all four realizations; the FID gaps were factors of 20.0 (PRAD), 6.9 (SKCM), 30.2 (LUSC) and 17.0 (NSCLC). Carried through pix2pixHD to RGB, the resulting images beat DCGAN-derived ones by 6.36× on H&E and 1.42× on IHC FID, and multi-label conditioning improved RGB FID by a further 14% over binary masks. No parameter counts are reported for either stage.

#Applications

The output is training data for pathology models — tissue segmentation, tumor detection, PD-L1 scoring — in cohorts where annotated slides are scarce or skewed. Because the label map is generated rather than inferred, every synthetic patch carries exact ground truth, and because sampling is unbounded a dataset can be enlarged past the size of the real cohort it came from. The multi-label variant makes composition controllable, so a set can be enriched for the PD-L1 expression patterns an immunotherapy classifier sees too rarely in practice.

#Impact

DEPAS established the noise-to-mask-to-image pipeline this group has built on since, and its coarse-mask successor benchmarks directly against it. Three limits are worth stating. Each cancer type and stain requires its own trained generator, with no transfer between organs demonstrated. The binary ground truth is a grayscale threshold rather than a pathologist annotation; only the multi-label PD-L1 masks are expert-annotated. Evaluation compares against held-out patches from the same cohorts, not an external one. No code, weights, or checkpoints have been released.