A GAN-based preprocessing approach for improved multiorgan segmentation

Multi-organ segmentation is a critical task in the pathological analysis of abdominal CT images. However, uneven brightness and incomplete organ edge contours often degrade image quality, severely affecting the accuracy of segmentation models. To address these issues, this paper proposes a novel preprocessing method based on a dual-input GAN architecture. The method employs two independent inputs: one input is the original CT images, while the other extracts style features from reference style images to achieve feature fusion. By incorporating a histogram discriminator, the method preprocesses the CT data to generate enhanced images that are better suited for multi-organ segmentation. This approach effectively preprocesses the dataset and improves segmentation performance. Experiments conducted on the BTCV (Beyond the Cranial Vault) dataset demonstrate the effectiveness of this method. When integrated with existing abdominal multi-organ segmentation algorithms, the method enhances performance, achieving an average Dice coefficient increase of approximately 0.164.