Source-linked AI summary
DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation
Hao Chen, Xiaojuan Qi, Lequan Yu, Pheng-Ann Heng
TL;DR
Manual gland annotation is laborious and unreliable, while histology variation and irregular malignant structures make automatic segmentation difficult. DCAN addresses this with multi-level contextual features and unified multi-task contour-aware learning, achieving the best results on Part A and surpassing other methods on Part B of the 2015 MICCAI challenge.
Problem
Manual gland annotation is time-consuming, effort-intensive, and limited in reproducibility, motivating efficient automatic segmentation for large-scale histopathological analysis.
Method
DCAN integrates multi-level contextual features with unified multi-task learning to jointly predict gland probability maps and contours without assuming glandular structure.
Results
The method achieved the best performance on testing Part A, while its non-contour-fused entry surpassed all other methods by a significant margin on Part B.
Takeaways & Limitations
Contour-aware multi-task learning helps separate clustered gland objects efficiently, supporting accurate gland segmentation on benchmark histology data.
Takeaways & Limitations
Irregular structures in malignant cases make gland segmentation more challenging, with lower performance on Part B than Part A.
Abstract
from arXiv · showhide
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.
1. Introduction
Gland morphology supports adenocarcinoma assessment, but accurate automated segmentation remains difficult because glands vary widely, touch, deform, and suffer preparation artifacts. DCAN addresses these challenges with multi-level contextual features and joint gland-contour learning, achieving strong benchmark performance.
- Motivation: Gland morphology is used to assess adenocarcinoma malignancy, making accurate segmentation essential for reliable tumor statistics.Manual annotation is labor-intensive and has limited reproducibility, motivating automatic methods for large-scale histopathological analysis.
- Challenges: Segmentation is challenged by variation across histologic grades and diseases, touching glands, malignant deformation, and tissue-preparation artifacts.These factors impede object separation and can invalidate assumptions based on regular glandular structure.
- Prior work: Earlier pixel-based and structure-based methods showed progress but often performed poorly on seriously deformed malignant glands.The remaining gap between algorithmic results and pathologist annotations motivated methods less dependent on hand-crafted features or structural priors.
- Proposed approach: DCAN uses multi-level contextual representations in an end-to-end network to generate gland probability maps efficiently.The approach avoids assuming a fixed glandular structure, supporting benign and malignant biopsy samples with different histopathological grades.
- Proposed approach: Joint learning of gland objects and contours enables simultaneous segmentation and separation of clustered glands.The contour-aware component provides complementary information for distinguishing touching objects, particularly in benign cases.
2. Method
DCAN extends fully convolutional segmentation with multi-level contextual features, auxiliary supervision, and jointly learned object and contour branches to separate touching glands.
- FCN with multi-level contextual features: Multi-level contextual features address gland-shape variation that a single receptive field cannot handle.Small receptive fields suit normal glands, whereas elongated malignant glands require larger contextual information.
- FCN with multi-level contextual features: Auxiliary classifiers C1-C3 strengthen deep-network training and alleviate vanishing gradients through additional supervision.Their weighted losses are reduced during training and omitted from the final simplified loss.
- Deep contour-aware network: Joint object-contour prediction addresses the ambiguity of touching glands that remains when segmentation relies only on gland-object likelihoods.Contour information complements object probabilities for separating clustered glands.
- Deep contour-aware network: The network jointly predicts gland-object and contour masks through complementary branches in an end-to-end multi-task framework.Separate upsampling branches infer object and contour probabilities, while shared hierarchical representations encode both types of information.
- Deep contour-aware network: The total loss combines L2 regularization with object and contour classification losses optimized over shared and task-specific parameters.The object and contour probabilities are trained against their respective ground-truth labels using standard back-propagation.
- Deep contour-aware network: Predicted object and contour probabilities are fused by thresholding, followed by smoothing, hole filling, small-area removal, and connected-component labeling.Both thresholds are set to 0.5 in the reported experiments.
- Transfer learning with rich feature hierarchies: Transfer learning initializes the downsampling path from a DeepLab model trained on PASCAL VOC 2012 to address scarce medical training data.The source dataset contains more than ten thousand pixel-annotated images, compared with a few hundred gland-segmentation images.
3. Experiments and results
The DCAN was evaluated on the Warwick-QU benchmark using challenge detection, segmentation, and shape-similarity criteria, with qualitative ablations and comparisons across benign and malignant cases. It achieved top overall performance, while contour fusion showed a malignant-case oversplitting trade-off.
- Experimental setup: The evaluation used the 2015 MICCAI Gland Segmentation Challenge dataset, with held-out testing Parts A and B and augmented training data.The dataset contains colorectal cancer histology images spanning benign to malignant grades, including poorly differentiated cases.
- Qualitative analysis: Qualitative ablations compared predictions without and with contour-aware information, displaying individual gland objects with different colors for benign and malignant cases.The comparison was intended to illustrate the contribution of complementary contour information.
- Evaluation criteria: The challenge measured gland detection, object-level segmentation, and shape similarity using F1 score, object-level Dice index, and Hausdorff distance.Ground-truth matching for detection used a 50% overlap criterion for true positives and false negatives.
- Quantitative comparison: CUMedVision1 without contour fusion surpassed all other methods on testing Part B, while CUMedVision2 achieved the best results on Part A.The results also indicate lower performance on Part B, which the authors associate with different data distributions and its predominance of malignant cases.
- Quantitative comparison: The deep contour-aware network achieved the best object-level Dice index on all testing data and CUMedVision2 obtained the smallest Hausdorff distance on Part A.The Part A Hausdorff distance was the only reported result below 50 pixels; CUMedVision1 was comparable to ExB1 on Part B for shape similarity.
- Overall results: The final ranking placed the network first out of 13 teams despite a contour-aware side effect in some malignant cases.The authors suggest selecting contour-aware fusion conditionally by first classifying images as benign or malignant.
4. Conclusions
The paper presents a deep contour-aware network that integrates multi-level contextual features and unified multi-task learning to segment glands accurately and separate clustered objects efficiently.
- DCAN integrates multi-level contextual features with unified multi-task learning for accurate gland segmentation.The framework harnesses complementary object and contour information rather than learning gland segmentation in isolation.