Calibration and Distraction Mining Network for Aortic True Lumen segmentation

Abstract

Automatic segmentation of aortic true lumen based on deep learning can save the time for diagnosis of aortic dissection. However, fuzzy boundary, small true lumen region, and high similarity usually leads to inaccurate prediction. To make better use of the details supplemented by the encoder to restore boundaries, we decompose the recovery of detail features in the decoder into two sub-processes: calibration and distraction mining. And we propose a novel calibration and distraction mining (CDM) module. It utilizes deep features to calibrate shallow features so that features are concentrated in the main region. Then, it leverages the distraction mining procedure to extract false-negative features as a supplement to calibrated features and recover details of the segmentation object. We construct CDM-Net and verify its performance on the Aorta-CT dataset (private dataset), it achieves the Dice similarity coefficient of 96.94% and the Jaccard index coefficient of 94.08%, which is the best compared with 10 latest methods. Similarly, we explore its robustness on three more public datasets, including ISIC 2018 dataset (skin lesion segmentation), the 2018 data science bowl dataset (nucleus segmentation), LUNA dataset (lung segmentation). Experimental results prove that our method produces competitive results on all three data sets. Through quantitative and qualitative research, the proposed CDM-Net has good performance and can process aortic slices with complex semantic features, additional experiments show that it has good robustness, and it has the potential to be applied and expanded conveniently.