An Inception-Multi-Scale-Attention U-Net for Breast Lesions Segmentation
in Ultrasound Images
Abstract
The study of segmenting breast lesions from ultrasound images is crucial
to breast cancer diagnosis and treatment. In this study, we propose the
Inception-Multi-Scale-Attention U-Net (IMSA-Net) method to address the
challenges associated with breast cancer imaging artifacts, tumor
morphological alterations, and blurred borders. The IMSA-Net method
incorporates the inception module into the downsampling feature
extraction section of the U-Net model. This can extract advanced
features from the input image and enhance the model’s capability to
express complex features. Furthermore, during the feature fusion stage,
a multi-scale attention structure is introduced to focus on capturing
detailed information during the segmentation process. This enhancement
contributes to improving the precision of the segmentation results. To
enhance the algorithm’s non-linear factor and mitigate the gradient
disappearance problem during training, the Gaussian Error Linear Unit
(GELU) activation function is employed, replacing the traditional
Rectified Linear Unit (ReLU) activation function. This modification
allows for better feature representation and overall model performance.
The experimental results demonstrate that the IMSA-Net model’s
evaluation metrics of mIoU, Dice, Acc, Precision, and Recall on the
breast ultrasound images dataset reached 74.18%, 77.98%, 96.86%,
89.23%, and 66.68%, respectively. These results outperform those of
the U-Net model by 2.41%, 2.81%, 0.09%, 0.06%, and 2.82%. Overall,
the IMSA-Net method tackles the difficulties related to breast lesion
segmentation, assisting in subsequent breast disease diagnosis.