Semantic-driven synthesis of histological images with controllable cellular distributions

Abstract

Digital pathology relies heavily on large, well-annotated datasets for training computational methods, but generating such datasets remains challenging due to the expertise required and inter-operator variability. We present SENSE (SEmantic Nuclear Synthesis Emulator), a novel framework for synthesizing realistic histological images with precise control over cellular distributions. Our approach introduces three key innovations: (1) A statistical modeling system that captures class-specific nuclear characteristics from expert annotations, enabling generation of diverse yet biologically plausible semantic content; (2) A hybrid ViT-Pix2Pix GAN architecture that effectively translates semantic maps into high-fidelity histological images; and (3) A modular design allowing independent control of cellular properties including type, count, and spatial distribution. Evaluation on the MoNuSAC dataset demonstrates that SENSE generates images matching the quality of real samples (MANIQA: 0.52 ± 0.03 vs 0.52 ± 0.04) while maintaining expert-verified biological plausibility. In segmentation tasks, augmenting training data with SENSE-generated images improved overall performance (DSC from 79.71 to 84.86) and dramatically enhanced detection of rare cell types, with neutrophil segmentation accuracy increasing from 40.18 to 78.71 DSC. This framework enables targeted dataset enhancement for computational pathology applications while offering new possibilities for educational and training scenarios requiring controlled tissue presentations.