3D-HLDM: Human-Guided Latent Diffusion Model to Improve Microvascular Invasion Prediction in Hepatocellular Carcinoma

Hyunho Shin, Nam-Joon Kim, Ji Hye Min, Seol Eui Lee, Ken Ying-Kai Liao, Hyuk-Jae Lee

Abstract: Microvascular invasion (MVI) is a critical risk factor for survival in patients with Hepatocellular Carcinoma. The presurgical prediction of MVI is clinically important and crucial for surgical and treatment planning. Although deep learning models have been employed to predict MVI using MRI, their performance has been limited because of data scarcity. To overcome this limitation, we propose a humanguided 3D Latent Diffusion Model (3D-HLDM) for generating a high-resolution synthetic MVI dataset. We examined our model using a clinical microvascular invasion (MVI)-MRI dataset with 475 cases provided by the Samsung Medical Center and various CNN-based prediction models. Consequently, we observed significant improvements in the performance of the prediction models when high-resolution synthetic images generated by 3D-HLDM were used.