Mixed 2D and 3D convolutional network with multi-scale context for lesion segmentation in breast DCE-MRI

Breast lesion segmentation in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is an essential step for breast cancer analysis. 2D networks have the advantages of low inference time and greater transfer-ability while 3D networks are superior in learning the contextual information in volumetric data. However, the performance of the 3D network is severely affected when MRI has a low-resolution ratio in the third dimension. In order to integrate the advantages of 2D and 3D networks, we propose a mixed 2D and 3D convolutional network with multi-scale context (M2D3D-MC) for lesion segmentation in breast DCE-MRI with a limited number of axial slices in the scans. Through serial 2D and 3D convolution, the mixed 2D and 3D convolution module has the ability to exploit the contexts between adjacent slices. And considering the diversity of shape and size for breast lesions, we introduce a multi-scale context extractor block consisting of atrous convolutions with different sampling rates to extract multi-scale image features. We justify the proposed method through extensive experiments on 90 MRI studies. Compared with both 2D and 3D networks, M2D3D-MC achieves the best performance with DSC, SEN, and PPV of 76.4%, 75.9%, and 82.4% respectively. A new paradigm is provided for breast lesion segmentation by combining 2D and 3D convolutions to exploit the contextual information. It demonstrates stronger performance in mixed 2D and 3D model given the limited number of axial slices. Our investigation also reveals that the multi-scale context block is effective for breast lesion segmentation.