Pyramid Deconvolution Net: Breast Cancer Detection Using Tissue and Cell Encoding Information
Dong Sui, Maozu Guo, Yue Zhang, Lei Zhang · 2020
Accurate diagnosis of breast cancer lesions from whole slide images (WSIs) is crucial for pathologist, since the results are associated with a certain status of breast cancer development. Until now, it is not yet possible to detect all of the cancerous areas in a WSIs. This limitation leads to mistakes or low detection precision in most of the approaches based on deep learning. In this paper, we propose an automatic cancer lesions detection approach using pyramid deconvolution network (PDN) for multi-level and multi-scale H&E stained breast pathological WSIs. Our workflow integrates tissue and cell level information for the cancerous region detection, and this is neglected by state-of-the-art methods. The high-level cancerous regions (macro) are obtained by deconvolution neuron network feature map from level 3 to level 9 in WSIs. The low-level cancer cells and region analysis pipeline is designed for detecting smaller region micro an ITCs from level 0. The results demonstrated that our workflow greatly improved the performance compared with those only using tissue level information.