Spatial Feature Conservation Networks (SFCNs) for Dilated Convolutions to Improve Breast Cancer Segmentation from DCE-MRI

Target delineation in the medical images can be utilized in lots of clinical applications, such as computer-aided diagnosis, prognosis, or radiation treatment planning. Deep learning has tremendously improved the performances of automated segmentation in a data-driven manner as compared with conventional machine learning models. In this work, we propose a spatial feature conservative design for feature extraction in deep neural networks. To avoid signal loss from sub-sampling of the max pooling operations, multi-scale dilated convolutions are applied to reach the large receptive field. Then, we propose a novel compensation module that prevents intrinsic signal loss from dilated convolution kernels. Furthermore, an adaptive combination method of the dilated convolution results is devised to enhance learning efficiency. The proposed model is validated on the delineation of breast cancer in DCE-MR images obtained from public dataset. The segmentation results clearly show that the proposed network model provides the most accurate delineation results of the breast cancers in the DCE-MR images. The proposed model can be applied to other clinical practice sensitive to spatial information loss.

Automated summary

In this study, a spatial feature conservative design for feature extraction in deep neural networks was proposed for target delineation in medical images. The model utilizes multi-scale dilated convolutions and a compensation module to enhance learning efficiency and prevent signal loss, achieving accurate delineation of breast cancer in the images.