Pathology generation leaves the 1024-pixel tile behind: coherent gigapixel slides, and synthetic samples lift few-shot classification by up to 18%
Synopsis
Histopathology image generation has stayed at the patch scale, producing tiles of at most 1024×1024 pixels, while clinical text reports and slide-level diagnostic labels act on the whole slide. BigCell trains a flow-matching transformer at 2048×2048 pixels on over 120,000 public whole-slide images and uses an inference-time sliding-window pipeline to generate coherent images up to 32,768×32,768 pixels, outperforming prior tile generators on tile-level and slide-level fidelity metrics. An adaptive scheduling strategy speeds generation by over three times with minimal quality degradation, and synthetic samples improve slide-level few-shot classification by up to 18%.
Interpretation
Histopathology generation moves from the patch scale to native gigapixel whole-slide scale, with coherence maintained across the entire slide. Existing histopathology generative models produce fixed-size tiles of at most 1024×1024 pixels, covering only a small fraction of a whole-slide image. Trained on over 120,000 public whole-slide images, BigCell outperforms prior tile generators on both tile-level and slide-level fidelity metrics.
Whole-slide generation comes from a flow-matching transformer trained at 2048×2048 pixels plus an inference-time sliding-window pipeline. A flow-matching transformer learns a continuous transformation from noise to image, and the sliding-window pipeline generates large images in pieces and stitches them into a full slide. Conditioning uses embeddings from pathology foundation models across multiple organ types, yielding coherent images up to 32,768×32,768 pixels.
To make gigapixel synthesis practical, adaptive scheduling directs computation toward important slide regions. Treating all regions uniformly makes whole-slide generation prohibitively expensive. The strategy speeds up image generation by over three times with minimal quality degradation.
Perspective
This report draws on summary-level material covering the stated training scale, generation resolution, speedup, and classification gain. It suits researchers and engineers working on histopathology image generation and slide-level data augmentation.
The material is a summary and does not give the specific fidelity metric values, the quantitative relation between speedup and quality degradation, or the exact classification setting behind the 18% gain.