Learning Adversarially Enhanced Heatmaps for Aorta Segmentation in CTA

Abstract

In this work, we propose a method to combine ADversarially enhanced HeatMaps (short for AD-HM) to segment the aorta from CTA (Computed Tomography Angiography). The intuition of the AD-HM is that heatmaps encompass rich information on locations of the targets. The positions of the aorta are relatively regular in CTA, thus training with heatmaps exploits the positional information to boost the segmentation results. The quality of heatmaps can be further enhanced with adversarial learning to refine the performance. The AD-HM can embed almost any state-of-the-art deep segmentation networks off the shelf. We collect 111 CTA volumes counting to 79082 slices to verify the effectiveness of our method. The training set is constituted of 104 volumes drawn from the dataset accounting to 74000 slices. The remaining 5082 slices from 7 CTA samples are reserved for validating the algorithm and the results are reported on the validation set. Our experiments with 7 state-of-the-art deep segmentation networks demonstrate the effectiveness of our method. The absolute improvement on IOU(Intersection-over-Union) of the aorta from the 7 models is 1.77% on average, with minimum improvement of 0.8% (UNet: 86.5%−> 87.3%) and maximum improvement of 3.4% (SegNet: 83.8%−> 87.2%).