Multi-scale context-aware networks for quantitative assessment of colorectal liver metastases

摘要

Colorectal Liver Metastases is the main cause of death in patients with colorectal cancer. Accurate histopathological assessment of the tumour regression - pathological response - in surgically resected liver matastases after preoperative chemotherapy is of paramount importance for prognostic stratification and subsequent planning of treatment. In this paper, we propose two multi-scale deep neural networks to address the absence of contextual information when using the patch-based methods for segmentation tasks in histopathological image analysis. The proposed networks are capable of integrating the texture features from a high magnification level and the contextual information from a low magnification level to achieve more accurate results. Extensive experiments have been performed on a dataset of whole slide scans of resected colorectal liver metastases annotated by experts. Our results demonstrate that the proposed multi-scale networks outperform the networks trained on any single magnification level alone.