SPAGHETTI

A phase-contrast image of a living cell and an H&E section of a tumour both encode morphology, but only one of them has a decade of pretrained models sitting behind it. Cellpose, StarDist, UNI, H-Optimus-0 and Phikon-v2 were all built on stained tissue, because that is where the large public corpora are. Phase-contrast microscopy (PCM) is the workhorse of live-cell culture precisely because it needs no stain, and that same absence of labels is why no comparable corpus exists to train on. SPAGHETTI inverts the usual response of gathering PCM images and fitting a bespoke model to them: rather than bringing models to the images, it moves the images to the models.

It is a cycle-consistent generative adversarial network trained on unpaired PCM and H&E images, so no field need ever be captured under both modalities. The addition that makes it usable is a structural-similarity (SSIM) term in the objective. Plain CycleGAN translation between imbalanced domains invents structure that was not in the input — phantom nuclei scattered across empty background — which would poison any downstream count or segmentation; penalising the generator whenever the output departs structurally from its input largely removes them. The authors are explicit about the result: a translated image "may not be biologically accurate as if cells were actually stained with H&E", and the method is a preprocessing step feeding other models, not a virtual stain for visual inspection.

SPAGHETTI comes from Gregory Schwartz's group at the Princess Margaret Cancer Centre, University Health Network, with the University of Toronto and the Vector Institute. The paper expands the name as Ssim-restrained PhAse contrast Gan to H&E TranslaTion of Images, the repository as Structural Phase Adaptation via Generative Histological Enhancement and Texture-preserving Translation Integration. The same lab's PENNE loads this checkpoint frozen as its front end.

#Key Features

• Structure-restrained translation: The SSIM loss anchors output luminance, contrast and structure to the input, cutting the nuclei an unmodified CycleGAN hallucinates.
• Unpaired training: The two image sets need not be equal in size or matched field-by-field, so any two public collections can serve as the domains.
• Drop-in preprocessing: The downstream segmenter or feature extractor runs unchanged and unaware, behind a pip-installable package, a CLI and a Dockerfile — for roughly 100 MB of extra GPU memory and 0.140 s per image.
• Optional adaptation to new imaging conditions: For images far from the training distribution — stained cells, a different microscope — self-supervised retraining on a few hundred target images plus a public H&E set restores performance.

#Technical Details

Both generators use a ResNet backbone with 9 residual blocks over 256x256 inputs; the two discriminators are PatchGAN classifiers operating on 16x16 patches. The deployed PCM-to-H&E generator holds about 11.4 M parameters, small enough to sit as a preprocessing tier under feature extractors hundreds of times larger. The objective sums adversarial, cycle-consistency, identity and SSIM terms, the latter three weighted 10.0, 5.0 and 10.0. Training used LIVECell for PCM (n = 5,326) and PanNuke for H&E (m = 7,904), an 80/20 split, batch size 16 and AdamW (learning rate 0.001, weight decay 0.01) for 100 epochs on 8 NVIDIA T4 GPUs.

On held-out LIVECell images, translation raised Cellpose's tissuenet_cp3 Dice score by 53.3% and StarDist's 2D_versatile_he by 95.0%, both past the raw-PCM baseline through Cellpose's cyto3. For cell-type annotation across eight cell lines, a random-forest classifier over UNI features reached 0.882 AUC with translation, against 0.870 for those same features without it. The sharpest gain came on an external C2C12 culture-media dataset absent from training, where H-Optimus-0 rose 18.1% to 0.910 AUC, past the 0.850 of the best untranslated arm. Attention maps trace the effect to translated images pulling a transformer's attention onto the cells.

#Applications

The natural home is a live-cell imaging pipeline stalled at "thousands of frames and no way to quantify them": outlining cells across a confluency time course, annotating cell types in mixed culture, classifying which growth-factor medium a well received, or regressing viability after a drug dose. Being a file-in, file-out step, it inserts ahead of existing analysis code and opens the pathology foundation-model ecosystem to labs that never make a stained slide.

#Impact

The wider argument is that a modality starved of training data can borrow another modality's models outright, provided the bridge is restrained enough not to invent evidence — far cheaper than pretraining a phase-contrast foundation model, and a recipe the authors point next at differential interference contrast and bright-field imaging. The limits are stated plainly: translated images lack the sectioning real H&E slides show and should not be read by eye, out-of-distribution inputs need a retraining pass that still requires a GPU, and the work remains a preprint awaiting peer review. Code is GPL-3.0 and the trained checkpoint ships in the repository, so the claims are open to independent test.