A Latent Diffuse Model for Synthetic Histopathology in Rare Cancers: Tackling Data Scarcity for AI Diagnostics

Extremely rare cancers such as sarcomas make AI-based diagnostics extremely difficult because of data scarcity. This work presents Sarco Diff, a novel latent diffusion model trained to generate high-resolution (1024×1024px) synthetic whole-slide histopathology of rare sarcoma subtypes. Using just 300 real images derived from The Cancer Genome Atlas (TCGA) and steered with a Low-Rank Adaptation (LoRA; Hu et al., 2021) on top, our model maintains diagnostically relevant features such as nuclear atypia and mitotic figures. In blinded assessments by five pathologists with board certifications, 41.7% of synthetic images were classified as real biopsies, respectfully, surpassing the performance for GAN-based alternatives (p=0.02). For a ResNet-50 classifier trained on both native and augmented data, detection of rare subtypes increased 25.3% using Sarco Diff-generated images (F1-score from 0.58→0.72), with the most pronounced improvements seen for individual subtypes where shown only <10 samples were available. For instance, an architecture with features yielding a FID score of score of 12.4 when validated, compared with 28.9 values for the state-of-the-art GANs. This foundational work establishes a novel approach to addressing data imbalance in computational pathology, by minimizing the reliance on rare tumour specimens while preserving diagnostic fidelity. Our method facilitates the generation of high-quality AI models for ultra-rare cancers, and can be adapted to other data-scarce medical imaging contexts.