Abstract: Predicting longitudinal changes in liver fibrosis is of great significance for guiding treatment decisions and improving patient prognosis. However, the diffuse distribution and spatiotemporal heterogeneity of fibrotic lesions, along with the subtle and inconsistent imaging differences between adjacent stages, pose significant challenges for accurate non-invasive assessment of disease progression. To address these issues, this paper proposes a Siamese-structured Conditional Gaussian Mixture Network (CGMNet) . First, a Two-Stage Cross Attention (TSCA) module is designed to promote feature interaction and fusion between baseline and follow-up images, thereby enhancing the model’s sensitivity to cross-temporal discriminative regions. Furthermore, to capture subtle inter-stage variations and alleviate feature distribution overlap, a Conditional Gaussian Mixture (CGM) Loss based on Gaussian mixture distribution is proposed. By using the staging labels of baseline images as conditions, this loss increases inter-class feature separation and improves intra-class feature compactness. The proposed method is evaluated on a self-constructed paired magnetic resonance imaging dataset of liver fibrosis patients. Experimental results show that CGMNet outperforms existing mainstream baseline models on multiple evaluation metrics, and ablation studies further confirm the individual contribution of each module to overall performance improvement.