Mt-UcGAN: Multi-task Uncertainty-Constrained GAN for Joint Segmentation, Quantification and Uncertainty Estimation of Renal Tumors on CT

Abstract

The segmentation of renal tumor, quantification of tumor indices (i.e., the center point coordinates, diameter, circumference, and cross-sectional area) and uncertainty estimation of segmentation are the key processes for clinical tumor disease diagnosis. However, these tasks have been studied independently so far. Because segmentation and quantification tasks have different optimization types, representing two different tasks as a unified optimization framework is a severe challenge. In this paper, we propose a unified framework (i.e., Mt-UcGAN: multi-task uncertainty-constrained generative adversarial network) for joint segmentation, quantification, and uncertainty estimation of renal tumors on CT. Mt-UcGAN includes a multitasking integrated generator (MtIG) and an uncertainty-constrained discriminator (UcD). MtIG achieves multi-task joint learning by novelly merging skip connections and Monte Carlo sampling. UCD guides the learning of segmentation and quantification networks by innovatively feeding prior information with high uncertainty constraints. Mt-UcGAN effectively corrects tumor prediction errors and improves network performance through continuous adversarial learning and alternate training. Experiments are performed on CT of 113 renal tumor patients. The dice coefficient of Mt-UcGAN is 92.1%, and the R2 coefficient of tumor circumference is 0.9513. The results show that this method has great potential to be extended to other medical image analysis tasks and clinical application value.