HL-FCN: Hybrid loss guided FCN for colorectal cancer segmentation
- 1 April 2018
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 195-198
Abstract
Colorectal cancer is among the leading cause of cancer-related mortalities. The cancerous regions are conventionally delineated from 3D magnetic resonance images in a voxel-wise way for radiotherapy. To ease the manual labeling procedure, which is laborious and time-consuming, automatic segmentation methods are highly demanded in clinical practice. However, it is a challenging task due to class imbalance and low-contrast appearance of cancerous regions, as well as the hard mimics from complex peritumoral areas. In this paper, we propose a volume-to-volume fully convolutional network architecture effectively trained with hybrid loss, referred as HL-FCN, to automatically segment colorectal cancer regions. Specifically, a novel Dice-based hybrid loss is designed under a multi-task learning framework to tackle the class-imbalance issue and hence improve the discrimination capability. Furthermore, a multi-resolution model ensemble strategy is developed to suppress false positives while preserving boundary details. Our method has been extensively validated on 64 cancerous cases using four-fold cross-validation, outperforming state-of-the-art methods by a significant margin.Keywords
This publication has 6 references indexed in Scilit:
- VoxResNet: Deep voxelwise residual networks for brain segmentation from 3D MR imagesNeuroImage, 2018
- 3D deeply supervised network for automated segmentation of volumetric medical imagesMedical Image Analysis, 2017
- DCAN: Deep contour-aware networks for object instance segmentation from histology imagesMedical Image Analysis, 2017
- Cancer statistics, 2017CA: A Cancer Journal for Clinicians, 2017
- V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation2016
- Automatic Detection and Segmentation of Crohn's Disease Tissues From Abdominal MRIIEEE Transactions on Medical Imaging, 2013