Progressive Dual Priori Network for Generalized Breast Tumor Segmentation

TL;DR

This work tackles the cross-center generalization challenge in breast tumor segmentation from dynamic contrast-enhanced MRI by proposing PDPNet, a two-part framework combining a coarse segmentation–guided localization module with a dual prior segmentation network (DPKNet) that leverages weak semantic priors and cross-scale correlation priors. The model progressively refines tumor masks after cropping to tumor regions, improving performance on small, low-contrast, and irregular tumors without requiring target-domain data. Ablation studies validate the contributions of localization and dual priors, and multi-center experiments show PDPNet achieving superior DSC and HD95 compared to state-of-the-art methods, with robust performance across diverse cohorts. The approach offers practical benefits for multicenter breast cancer imaging and lays groundwork for extending to 3D/4D data and more sophisticated priors.

Abstract

To promote the generalization ability of breast tumor segmentation models, as well as to improve the segmentation performance for breast tumors with smaller size, low-contrast and irregular shape, we propose a progressive dual priori network (PDPNet) to segment breast tumors from dynamic enhanced magnetic resonance images (DCE-MRI) acquired at different centers. The PDPNet first cropped tumor regions with a coarse-segmentation based localization module, then the breast tumor mask was progressively refined by using the weak semantic priori and cross-scale correlation prior knowledge. To validate the effectiveness of PDPNet, we compared it with several state-of-the-art methods on multi-center datasets. The results showed that, comparing against the suboptimal method, the DSC and HD95 of PDPNet were improved at least by 5.13% and 7.58% respectively on multi-center test sets. In addition, through ablations, we demonstrated that the proposed localization module can decrease the influence of normal tissues and therefore improve the generalization ability of the model. The weak semantic priors allow focusing on tumor regions to avoid missing small tumors and low-contrast tumors. The cross-scale correlation priors are beneficial for promoting the shape-aware ability for irregular tumors. Thus integrating them in a unified framework improved the multi-center breast tumor segmentation performance. The source code and open data can be accessed at https://github.com/wangli100209/PDPNet.

Figures (12)

• Figure 1: Overall structure of PDPNet. It consists of a localization module which is used to limit the influence of the background information, and a dual prior knowledge based segmentation network (DPKNet) which is responsible for segmenting the breast tumors in a progressive manner.
• Figure 2: Dual priori knowledge based network (DPKNet). It has an encoder-decoder architecture, with a dual prior module (DPM) being used between the low-level encoded features $M_{s-1}$ and high-level decoded features $F_s, (s={3,4,5})$ to promote the segmentation performance.
• Figure 3: Visual comparisons among various methods on validation set (a) and multi-center test sets (b). The original images and the ground truth tumor masks are presented in the first and second columns, respectively. In each sub-figure, the ground-truth tumor regions are marked in blue, the predicted tumor regions are highlighted in red and the overlap regions between segmented results and ground truth are shown in yellow. For better visualization, the local segmented tumor details are zoomed in and shown in the bottom.
• Figure 4: Radar plots of average evaluation metrics obtained with different methods on both validation (a) and multi-center test sets (b). The number indicated in each legend represents the surface of area enclosed by three metrics. The bigger area indicates the better performance.
• Figure 5: The curves of average DSC for all the samples in five test cohorts obtained by different methods.