A pathology image segmentation framework based on deblurring and region proxy in medical decision-making system

The cell nuclei in pathology images provide clinicians with important tissue information for diagnosis. However, cell nucleus segmentation faces various challenges such as image blurring, blurred cell nucleus boundaries, noise, and holes. Existing segmentation methods are difficult to achieve the expected results in the field of cell nuclei. Based on this, this study proposes a framework for pathology image segmentation based on deblurring and region proxies (DRPVit) in medical decision-making systems. First, we employ an information-enhanced iterative filtering adaptive network (IE-IFANet) to improve the clarity of nucleus regions and boundaries. Then, considering the performance and complexity of the network, the RegProxy model is used for cell nucleus segmentation. The RegProxy model utilizes an agent to represent regions with homogeneous semantics, a transformer encoder to model the relationship between these regions, and finally a classifier for region prediction. Finally, we introduce a combined loss function consisting of a boundary loss and a tversky loss to balance and optimize the segmentation of edges and regions. In the post-processing stage, we eliminated the holes and noise in the predicted images and introduced paramedical tools to measure the cell nucleus size and morphological differences. The results of the experimental evaluation show that our method outperforms the comparative models with an improvement of 3.3%, 2.3%, and 2.1% for Intersection over Union, Dice similarity coefficient, and Recall, respectively.