![[2512.01292] Diffusion Model in Latent Space for Medical Image Segmentation Task](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2512.01292] Diffusion Model in Latent Space for Medical Image Segmentation Task
Summary
This article presents MedSegLatDiff, a novel diffusion model for efficient medical image segmentation that enhances interpretability by generating multiple plausible masks per image.
Why It Matters
Medical image segmentation is critical for accurate clinical diagnosis and treatment. Traditional methods often fail to capture uncertainty, while MedSegLatDiff offers a more reliable approach by generating diverse segmentation hypotheses, making it particularly valuable in clinical settings.
Key Takeaways
- MedSegLatDiff combines a variational autoencoder with a latent diffusion model for efficient segmentation.
- The model generates multiple plausible segmentation masks, improving interpretability and reliability.
- It replaces conventional loss functions to better preserve small structures in medical images.
- Evaluated on multiple datasets, it achieves state-of-the-art performance in segmentation metrics.
- This approach is particularly suited for clinical deployment due to its enhanced reliability.
Computer Science > Computer Vision and Pattern Recognition arXiv:2512.01292 (cs) [Submitted on 1 Dec 2025 (v1), last revised 26 Feb 2026 (this version, v3)] Title:Diffusion Model in Latent Space for Medical Image Segmentation Task Authors:Huynh Trinh Ngoc, Toan Nguyen Hai, Ba Luong Son, Long Tran Quoc View a PDF of the paper titled Diffusion Model in Latent Space for Medical Image Segmentation Task, by Huynh Trinh Ngoc and 3 other authors View PDF HTML (experimental) Abstract:Medical image segmentation is crucial for clinical diagnosis and treatment planning. Traditional methods typically produce a single segmentation mask, failing to capture inherent uncertainty. Recent generative models enable the creation of multiple plausible masks per image, mimicking the collaborative interpretation of several clinicians. However, these approaches remain computationally heavy. We propose MedSegLatDiff, a diffusion based framework that combines a variational autoencoder (VAE) with a latent diffusion model for efficient medical image segmentation. The VAE compresses the input into a low dimensional latent space, reducing noise and accelerating training, while the diffusion process operates directly in this compact representation. We further replace the conventional MSE loss with weighted cross entropy in the VAE mask reconstruction path to better preserve tiny structures such as small nodules. MedSegLatDiff is evaluated on ISIC-2018 (skin lesions), CVC-Clinic (polyps), and LIDC-IDRI (l...
