V‐UNet: Medical Image Segmentation Based on Variational Attention Mechanism

ABSTRACT Accurate medical image segmentation plays a crucial role in improving the precision of computer‐aided diagnosis. However, complex boundary shapes, low contrast and blurred anatomical structures make fine‐grained segmentation a challenging task. Variational Bayesian inference quantifies uncertainty through probability distributions and can construct robust probabilistic models for the boundaries of ambiguous organs and tissues. In this paper, we apply variational Bayesian inference to medical image segmentation and propose variational attention to model the uncertainty of low‐contrast and blurry tissue and organ boundaries. This enhances the model's ability to perceive segmentation boundaries, improving robustness and segmentation accuracy. Variational attention first estimates the parameters of the probability distribution of latent representations based on input features. Then, it samples latent representations from the learnt distribution to generate attention weights that optimise the interaction between global features and ambiguous boundaries. We integrate variational attention into the U‐Net model by replacing its skip connections, constructing a multi‐scale variational attention segmentation model (V‐UNet). Experiments on the ISBI 2012 and MoNuSeg 2018 datasets show that our method achieves Dice scores of 95.89% and 82.18%, respectively. Moreover, we integrate V‐UNet into the Mask R‐CNN framework by replacing the FPN feature extraction head and propose a two‐stage segmentation method. Compared to the original Mask R‐CNN, our method improves the Dice score by 0.81%, mAP by 8.06% and F1 score by 0.51%.