HISPACE: Histological Image Synthesis with Pattern And Content Engine

The growing integration of Digital Pathology (DP) within histological analysis has opened new perspectives in biomedical imaging and data-driven research. Whole Slide Imaging (WSI) technologies enable large-scale digitalization of tissue samples, allowing visualization, storage and computational processing. However, the development of reliable data-driven methods remains limited by the lack of annotated dataset and especially due to the high inter-operator variability in the labelling process performed by pathologists. To address these challenges, this thesis proposes to invert the traditional paradigm by introducing a controlled generation framework for synthetic histological images with user-defined semantic content. The proposed system called HISPACE (Histological Image Synthesis with Pattern And Content Engine) is based on a hybrid pipeline, where semantic content generation is driven by explicit mathematical modelling, while the subsequent translation into photorealistic images is achieved through a GAN model. In the first stage, each cell contour is reconstructed using Fourier Descriptors (FDs), ensuring morphological coherence with real samples, while the overall tissue layout is modelled through Kernel Density Estimation (KDE) to reproduce realistic spatial patterns derived from real data distributions. The model allows direct user control over key semantic parameters – the number of cells, tissue density and spatial configuration – enabling the creation of realistic and customizable histological images. The pipeline was developed and validated using two annotated datasets, MoNuSAC and a private dataset derived from liver WSI samples. Quantitative evaluation was conducted through visual similarity metrics (e.g. FID, no-reference quality measures) and by assessing the performances over a steatosis segmentation network. Additionally, a paired visual comparison and expert evaluation were performed to determine the realism of generated images.