Abstract
Medical image segmentation often involves inherent uncertainty due to inter observer variability. In this case, a single deterministic mask obtained by conventional segmentation networks, such as U-Net, cannot capture the distribution of plausible expert annotations, risking missed clinically relevant variants. In order to enable uncertainty quantification and reflect inter expert variability, probabilistic models like Probabilistic U-Net are used, which perform aleatoric or ambiguous segmentation where a latent space is sampled to generate multiple segmentation masks. However, the common use of a conditioned unimodal posterior in these models fails to represent true multimodality, leading to mode bias and limited diversity. To address these limitations, a multi-level Probabilistic U-Net augmented with normalizing flows is proposed to enhance the expressiveness of the latent distribution. The multi-level design induces multiple latent distributions in separate levels of U-Net, enabling more diverse sampling, while the flow module transforms the posterior to add data required modes and expand representational capacity, thereby enriching the expressiveness of the distributions. The proposed flow incorporated multi-level network enables a more flexible and powerful distribution, thereby enhancing the model’s ability to generate high fidelity segmentation masks. Extensive experiments on some publicly available datasets with multiple expert annotations per image demonstrate that the proposed model reduces generalized energy distance (GED), preserves clinically meaningful diversity and sharpens boundary fidelity, with latent grid analyses indicating fuller mode coverage and fewer artifacts. Collectively, these results indicate that the proposed framework advances accuracy, robustness, and clinical reliability for aleatoric, uncertainty aware medical image segmentation.