A hybrid hemodynamic knowledge-powered and feature reconstruction-guided scheme for breast cancer segmentation based on DCE-MRI
Tianxu Lv,Youqing Wu,Yihang Wang,Yuan Liu,Lihua Li,Chuxia Deng,Xiang Pan
DOI: https://doi.org/10.1016/j.media.2022.102572
Abstract:Automatically and accurately annotating tumor in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), which provides a noninvasive in vivo method to evaluate tumor vasculature architectures based on contrast accumulation and washout, is a crucial step in computer-aided breast cancer diagnosis and treatment. However, it remains challenging due to the varying sizes, shapes, appearances and densities of tumors caused by the high heterogeneity of breast cancer, and the high dimensionality and ill-posed artifacts of DCE-MRI. In this paper, we propose a hybrid hemodynamic knowledge-powered and feature reconstruction-guided scheme that integrates pharmacokinetics prior and feature refinement to generate sufficiently adequate features in DCE-MRI for breast cancer segmentation. The pharmacokinetics prior expressed by time intensity curve (TIC) is incorporated into the scheme through objective function called dynamic contrast-enhanced prior (DCP) loss. It contains contrast agent kinetic heterogeneity prior knowledge, which is important to optimize our model parameters. Besides, we design a spatial fusion module (SFM) embedded in the scheme to exploit intra-slices spatial structural correlations, and deploy a spatial-kinetic fusion module (SKFM) to effectively leverage the complementary information extracted from spatial-kinetic space. Furthermore, considering that low spatial resolution often leads to poor image quality in DCE-MRI, we integrate a reconstruction autoencoder into the scheme to refine feature maps in an unsupervised manner. We conduct extensive experiments to validate the proposed method and show that our approach can outperform recent state-of-the-art segmentation methods on breast cancer DCE-MRI dataset. Moreover, to explore the generalization for other segmentation tasks on dynamic imaging, we also extend the proposed method to brain segmentation in DSC-MRI sequence. Our source code will be released on https://github.com/AI-medical-diagnosis-team-of-JNU/DCEDuDoFNet.