Multilevel perception boundary-guided network for breast lesion segmentation in ultrasound images

Automatic segmentation of breast tumors from the ultrasound images is essential for the subsequent clinical diagnosis and treatment plan. Although the existing deep learning-based methods have achieved considerable progress in automatic segmentation of breast tumors, their performance on tumors with similar intensity to the normal tissues is still not satisfactory, especially for the tumor boundaries.

To accurately segment the non-enhanced lesions with more accurate boundaries, a novel multilevel perception boundary-guided network (PBNet) is proposed to segment breast tumors from ultrasound images.

PBNet consists of a multilevel global perception module (MGPM) and a boundary guided module (BGM). MGPM models long-range spatial dependencies by fusing both intra- and inter-level semantic information to enhance tumor recognition. In BGM, the tumor boundaries are extracted from the high-level semantic maps using the dilation and erosion effects of max pooling; such boundaries are then used to guide the fusion of low- and high-level features. Additionally, a multi-level boundary-enhanced segmentation (BS) loss is introduced to improve boundary segmentation performance. To evaluate the effectiveness of the proposed method, we compared it with state-of-the-art methods on two datasets, one publicly available datasets BUSI containing 780 images and one in-house dataset containing 995 images. To verify the robustness of each method, a 5-fold cross-validation method was used to train and test the models. Dice score (Dice), Jaccard coefficients (Jac), Hausdorff Distance (HD), Sensitivity (Sen), and specificity(Spe) were used to evaluate the segmentation performance quantitatively. The Wilcoxon test and Benjamini-Hochberg false discovery rate (FDR) multi-comparison correction were then performed to assess whether the proposed method presents statistically significant performance ( $p\le 0.05$ ) difference comparing with existing methods. In addition, to comprehensively demonstrate the difference between different methods, the Cohen's d effect size and compound p-value (c-Pvalue) obtained with Fisher's method were also calculated.

On the BUSI dataset, the mean Dice and Sen of PBNet was increased by 0.93% ( $p\le 0.01$ ) and 1.42% ( $p\le 0.05$ ), respectively, comparing against the corresponding suboptimal methods. On the in-house dataset, PBNet improved Dice, Jac and Spe by approximately 0.86% ( $p\le 0.01$ ), 1.42% ( $p\le 0.01$ ), and 0.1%, respectively, and reduced HD by 1.7% ( $p\le 0.01$ ) compared to the sub-optimal model. Comprehensively, in terms of all the evaluation metics, the performance of the proposed method significantly (c-Pvalue $\le 0.05$ ) outperformed the others but the effect size was smaller than 0.2. Ablation results confirmed that MGPM is effective in distinguishing non-enhanced tumors, while BGM and BS loss are beneficial for refining tumor segmentation contours.

The proposed PBNet allows us to segment the non-enhanced breast lesions from ultrasound images with more accurate boundaries, which provides a valuable means for the subsequent clinical applications.