Generating Non-Melanoma Skin Cancer Histopathology Images Using Latent Diffusion Representation Learning

The generation of skin cancer histopathology images plays a crucial role in addressing the limitations of small and unbalanced datasets, which helps improve skin cancer diagnosis. Acquiring these im-ages is difficult because histopathology Whole Slide Images (WSIs) are extremely large, often reaching gigapixel sizes and are costly to produce. Recently, generative models have emerged as powerful tools capable of synthesizing high-quality medical images, with diffusion models standing out for their exceptional ability to capture fine details and complex struc-tures. However, diffusion models face a trade-off between image quality and inference time, because they operate directly in pixel space, they demand substantial computational resources, resulting in longer infer-ence times and increased model complexity. To address this challenge, we introduce a customized Latent Diffusion Model framework named PathoLDM. This framework compresses pixel-domain data into a lower-dimensional latent space using an autoencoder, effectively preserving tex-ture, cellular and tissue structures, as well as semantic information in skin cancer WSI images, while significantly reducing computational overhead. Our framework retains essential histopathological characteristics such as nuclear morphology, melanin distribution, and tumor boundary integrity, enabling high-fidelity image synthesis and robust reconstruction of miss-ing or degraded regions, without compromising the pathological authen-ticity of the textures. We evaluated our framework through extensive ex-periments on a publicly available dataset: the Non-Melanoma Skin Can-cer (NMSC) Histopathology dataset. Our results show that PathoLDM consistently outperformed existing state-of-the-art models in both gen-erative quality and computational efficiency.

N/A

CBM 2025: Computational Biomechanics for Medicine Workshop

N/A

2195-2728

N/A

N/A

10