Hybrid receptive field U-Net for lesion segmentation in breast ultrasound images

Numerous neural networks based on the U-Net architecture have been developed for the segmentation of breast ultrasound images. However, the accuracy of such segmentation tasks is often compromised by the complex and variable shapes of tumors, the indistinct boundaries of lesion areas, and background noise.

In this study, we develop a hybrid receptive field U-Net (HRF U-Net) to improve lesion segmentation in breast ultrasound images.

We design an innovative hybrid receptive field module (HRFM) to replace conventional convolutional layers, integrating deformable convolutions that expand the receptive field and enhance the model's ability to capture shape and boundary features. We systematically analyze the differential feature extraction capabilities of deformable convolutions across network layers, combining them with standard convolutions, dilated convolutions, and max pooling. This configuration enables HRF U-Net to achieve a broader and more adaptive receptive field, enhancing its effectiveness in handling the complexities of breast lesion segmentation. Additionally, we introduce a large-kernel attention module (LKAM) within the skip connections, which expands the receptive field and supports adaptive feature selection, capturing long-range dependencies within the convolutional attention mechanism. This novel approach enables more precise feature extraction, effectively mitigates boundary noise during training, and substantially improves the model's segmentation performance. We used three publicly available datasets to conduct extensive experiments, including ablation studies, comparative analyses, robustness evaluations, and external validation. Datasets A and B were divided into training and validation sets for four-fold cross-validation, while Dataset C was used as the test set for external validation. We selected five widely used image segmentation metrics for validation, namely pixel accuracy, precision, recall, Jaccard index, and Dice coefficient. In addition, the statistical significance was evaluated using the paired Student's t-test with Holm-Bonferroni correction $(\alpha =0.05)$ .

Experimental results on three public datasets demonstrate that the proposed HRF U-Net substantially improves the efficacy of breast ultrasound image segmentation, outperforming several state-of-the-art works. Specifically, in the comparative experiments, HRF U-Net achieved scores of 96.52, 85.03, 83.19, 73.15, and 81.34 for the five metrics on Dataset A. On Dataset B, the scores reached 98.61, 90.22, 83.99, 74.50, and 83.64 . The effect size Cohen's d values for HRF U-Net compared to several typical and advanced networks all exceed 0.32, demonstrating a significant improvement in segmentation performance. In external validation, the metric scores reached 97.02, 84.35, 93.94, 77.74, and 86.86 .

The HRF U-Net proposed in this paper improves the segmentation accuracy of breast ultrasound images.