BigCell: Generating Gigapixel Whole-Slide Images

BigCell. Generating Gigapixel Whole-Slide Images. By Srikar Yellapragada and twelve others, from Stony Brook University, Argonne National Laboratory, and three other institutions. Published on the ninth of October. No code repository is linked to it.

Whole-slide images are digital scans of histopathology tissue at gigapixel scale, and they serve as the primary basis for cancer diagnosis. Existing generative models for histopathology work at the scale of a patch. They produce fixed-size tiles of at most one thousand and twenty-four by one thousand and twenty-four pixels, which cover only a small fraction of a whole slide.

Yet the supervision that exists at scale lives at the level of the whole slide. Clinical text reports describe whole slides. Diagnostic labels are attached to whole slides. So the generative models work on small pieces, while the labels that are plentiful describe the full image. The two sit at different sizes.

BigCell closes that gap from the generation side. It is presented as the first approach to generate coherent whole-slide histopathology images at native gigapixel resolution.

It trains a flow-matching transformer at two thousand and forty-eight by two thousand and forty-eight pixels. The training data is over one hundred and twenty thousand publicly available whole-slide images, spanning multiple organ types. Generation is conditioned on embeddings from pathology foundation models. At inference, a sliding-window pipeline extends the output to coherent images of up to thirty-two thousand seven hundred and sixty-eight pixels on each side.

Making synthesis at this size practical takes one more piece. An adaptive scheduling strategy prioritizes the important regions of a slide, which speeds up generation by more than three times with minimal loss of quality. Separately, the authors train a second model that produces dense conditioning grids directly from pathology text reports. That gives what they call the first text-to-image synthesis at eight K resolution.

BigCell outperforms prior tile generators on fidelity metrics at both the tile level and the slide level. The synthetic samples are also useful downstream. Adding them to training improves few-shot classification at the slide level by up to eighteen percent, compared with baselines trained on real data.

The move is to generate at the scale where the labels already are. Once a model can produce a whole slide, the reports and diagnoses that exist in quantity at that scale become usable as conditioning. The synthetic slides then feed back into classification, where they add to what real data alone provides.