[Paper Review] DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation

Hao Chen, Xiaojuan Qi+2 more|arXiv (Cornell University)|Apr 10, 2016

AI in cancer detection44 references130 citations

TL;DR

DCAN introduces a deep contour-aware network with multi-level contextual features and a contour supervision branch in a unified multi-task framework to accurately segment glands and separate touching glands in histology images, achieving top performance at MICCAI 2015.

ABSTRACT

The morphology of glands has been used routinely by pathologists to assess the malignancy degree of adenocarcinomas. Accurate segmentation of glands from histology images is a crucial step to obtain reliable morphological statistics for quantitative diagnosis. In this paper, we proposed an efficient deep contour-aware network (DCAN) to solve this challenging problem under a unified multi-task learning framework. In the proposed network, multi-level contextual features from the hierarchical architecture are explored with auxiliary supervision for accurate gland segmentation. When incorporated with multi-task regularization during the training, the discriminative capability of intermediate features can be further improved. Moreover, our network can not only output accurate probability maps of glands, but also depict clear contours simultaneously for separating clustered objects, which further boosts the gland segmentation performance. This unified framework can be efficient when applied to large-scale histopathological data without resorting to additional steps to generate contours based on low-level cues for post-separating. Our method won the 2015 MICCAI Gland Segmentation Challenge out of 13 competitive teams, surpassing all the other methods by a significant margin.

Motivation & Objective

• Motivate automatic, accurate gland segmentation to assist histopathological analysis across benign and malignant cases.
• Develop an end-to-end FCN-based framework that leverages multi-level contextual features for robust gland delineation.
• Incorporate contour information via a deep contour-aware multi-task network to separate touching gland objects.
• Enable transfer learning to mitigate limited medical training data.
• Demonstrate efficiency suitable for large-scale histopathology workflows.

Proposed method

• Extend fully convolutional networks (FCN) to fuse multi-level contextual features from different receptive fields for pixel-wise gland probability maps.
• Add auxiliary supervision to stabilize training and improve feature discriminativeness across deep layers.
• Introduce a deep contour-aware network with two output branches for gland objects and contours, sharing the downsampling path and training jointly in a multi-task loss.
• Define an end-to-end loss that combines regularization with learned per-pixel gland and contour predictions.
• Fuse gland object and contour probability maps to obtain final gland segmentation with thresholds t_o and t_c and post-processing steps.
• Apply transfer learning by initializing the downsamping path with DeepLab/PASCAL-VOC weights and fine-tuning on the gland dataset.

Experimental results

Research questions

• RQ1 Can multi-level contextual features improve gland segmentation accuracy across varied histology grades?
• RQ2 Does incorporating a contour branch help separate touching or clustered glands more effectively than object-only segmentation?
• RQ3 Does a unified multi-task framework yield faster, more robust gland segmentation suitable for large-scale data?
• RQ4 What is the impact of transfer learning from natural image datasets on gland segmentation performance?

Key findings

• The deep contour-aware network (DCAN) achieves superior detection and segmentation performance on the 2015 MICCAI Gland Segmentation Challenge Part A and Part B.
• Contour-aware fusion (CUMedVision2) yielded best detection on Part A and competitive results on Part B, achieving top overall ranking among 13 teams.
• Object-level Dice and Hausdorff metrics show DCAN with contour guidance outperforms competing methods, with the smallest Hausdorff distance reported for Part A.
• Inference is fast, averaging about 1.5 seconds per 755×522 image, enabling practical use on large-scale histology data.
• Training time is about four hours on a Titan X GPU, illustrating efficient optimization with transfer learning.
• Ablation analysis confirms that contour information improves separation of touching glands, particularly in benign cases.