CUVAE: Strengthening Latent Representations in Skip-Connection VAEs for High-Fidelity Medical Image Reconstruction

Variational Autoencoders (VAEs) with skip-connections often suffer from posterior collapse, where the latent space becomes disorganized as the decoder bypasses the bottleneck to utilize high-frequency spatial information. This is particularly detrimental in clinical diagnostics, where a structured latent manifold is essential for feature disentanglement. In this paper, we propose the Constrained Unfolding Variational Autoencoder (CUVAE), a novel architecture that utilizes weighted skip-connections and batch-normalized latent constraints to preserve semantic organization. We validated our approach using the Chest X-ray (Pneumonia) and BraTS 2020 (Brain Tumor) datasets. Qualitative assessment via refined t-SNE visualization ($Perplexity=80$, $Exaggeration=18$) demonstrates that while baseline models exhibit significant cluster overlap, the proposed CUVAE yields a distinct manifold topology with clear separation between healthy tissue and pathological etiologies (Bacterial vs. Viral pneumonia; Tumor Core vs. Edema). Quantitative results indicate a significant improvement in latent semantic separation, suggesting that CUVAE effectively encodes clinically relevant biomarkers without sacrificing reconstruction fidelity.