Source-linked AI summary
CIA-Net: Robust Nuclei Instance Segmentation with Contour-aware Information Aggregation
Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng
TL;DR
CIA-Net addresses over- and under-segmentation from clustered nuclei, morphological variation, and noisy annotations that limit robust instance segmentation. It aggregates information between nuclei and contour decoders and uses a smooth truncated loss; on the 2018 MICCAI challenge, it surpassed the other 35 teams.
Problem
Nuclei instance segmentation is challenged by clusters, indistinguishable boundaries, subjective mislabeling, and variation across organs, which hinder robust analysis and generalization.
Method
CIA-Net uses hierarchical, bidirectional multi-level feature aggregation between task-specific nuclei and contour decoders, with a smooth truncated loss that reduces outlier interference.
Results
CIA-Net surpassed all other 35 teams in the 2018 MICCAI Multi-Organ-Nuclei-Segmentation challenge.
Takeaways & Limitations
The method supports robust nuclei instance segmentation under clustered nuclei and labeling noise, including evaluation on unseen organs.
Abstract
from arXiv · showhide
Accurate segmenting nuclei instances is a crucial step in computer-aided image analysis to extract rich features for cellular estimation and following diagnosis as well as treatment. While it still remains challenging because the wide existence of nuclei clusters, along with the large morphological variances among different organs make nuclei instance segmentation susceptible to over-/under-segmentation. Additionally, the inevitably subjective annotating and mislabeling prevent the network learning from reliable samples and eventually reduce the generalization capability for robustly segmenting unseen organ nuclei. To address these issues, we propose a novel deep neural network, namely Contour-aware Informative Aggregation Network (CIA-Net) with multi-level information aggregation module between two task-specific decoders. Rather than independent decoders, it leverages the merit of spatial and texture dependencies between nuclei and contour by bi-directionally aggregating task-specific features. Furthermore, we proposed a novel smooth truncated loss that modulates losses to reduce the perturbation from outliers. Consequently, the network can focus on learning from reliable and informative samples, which inherently improves the generalization capability. Experiments on the 2018 MICCAI challenge of Multi-Organ-Nuclei-Segmentation validated the effectiveness of our proposed method, surpassing all the other 35 competitive teams by a significant margin.
1 Introduction
Nuclei instance segmentation is important for cellular estimation and pathology, but clusters, ambiguous boundaries, noisy annotations, and organ-specific variation make robust automation difficult. CIA-Net addresses these challenges by aggregating contour and nuclei information while reducing outlier influence during training.
- Nuclei instance segmentation captures location, density, and morphology features needed for tumor diagnosis and treatment procedures.
- Clusters and occlusions can cause over- or under-segmentation, impairing accurate morphological measurements of nuclei instances.
- Blurred borders and inconsistent staining produce indistinguishable instances, subjective annotations, and mislabeling that challenge robust results.
- Noisy labels can overwhelm loss calculations and dominate gradients, preventing learning from informative samples.
- CIA-Net uses bidirectional feature aggregation between nuclei and contour decoders, while smooth truncated loss reduces outlier perturbation and supports generalization.The method achieved first place in the 2018 MICCAI Challenge on Multi-Organ Nuclei Segmentation.
2 Method
CIA-Net combines a densely connected encoder with pyramidal refinement and two task-specific decoders that exchange nuclei and contour information. Its Smooth Truncated Loss limits outlier influence while retaining Soft Dice Loss for contour shape learning.
- Network architecture: CIA-Net uses a densely connected encoder and two task-specific information-aggregated decoders for hierarchical multi-level refinement.Dense connectivity reuses features through direct concatenation, while the decoders aggregate task-specific features rather than operating independently.
- Network architecture: Pyramidal lateral connections merge encoder features at 1/2, 1/4, and 1/8 resolution with upsampled decoder features.Earlier layers contribute localization and texture information, while deeper features provide stronger semantics for detail refinement.
- Information aggregation: The Information Aggregation Module bidirectionally exchanges spatial and contextual cues between nuclei and contour branches to refine both subtasks.The module is motivated by the high spatial and contextual relevance between contour regions and their corresponding nuclei.
- Robust loss: Smooth Truncated Loss clips low-probability outliers below γ while preserving the loss for other samples, reducing gradient domination by noisy labels.The smoothed formulation replaces the abrupt truncation with a quadratic function matching the negative log-likelihood value and derivative at γ.
- Robust loss: CIA-Net combines Smooth Truncated Loss for nuclei prediction with Soft Dice Loss for contour prediction and weight decay in the total objective.Contour prediction is treated differently because its high-loss regions are primarily inliers associated with irregularly curved shapes rather than outliers.
3 Experimental Results
Experiments on seen and unseen organs evaluate CIA-Net’s contour-aware aggregation and Smooth Truncated loss. The method improves instance segmentation, produces clearer heatmaps across γ values, and surpasses competing challenge entries.
- Dataset and evaluation: The MoNuSeg dataset contains 30 whole-slide images spanning seven organs, with Test2 comprising unseen bladder, colon, and stomach organs.Train, Test1, and Test2 contain 14, 8, and 6 images, respectively.
- Architecture comparison: CNN-based methods substantially outperform conventional thresholding and watershed approaches across the reported evaluation criteria.The comparison includes Cell Profiler, Fiji, CNN3, PA-Net, BES-Net, CIA-Net without IAM, and CIA-Net.
- Architecture comparison: CIA-Net consistently achieves the best AJI compared with BES-Net and CIA-Net without IAM, supporting bidirectional contour–nuclei feature refinement.The Information Aggregation Module leverages spatial and contextual relevance between the two tasks.
- Loss-function comparison: LST improves AJI on unseen-organ Test2 by nearly 2.5%, while LT improves it by 1% relative to the first two loss-function baselines.LST also outperforms LT on Test2 and is less sensitive to γ selection.
- Loss-function comparison: As γ increases, LST heatmaps become more concrete with fewer uncertain areas, whereas overly large γ increases under-segmentation risk.The text attributes this risk to suppressing informative low-probability inlier samples, particularly near boundaries.
- Challenge and qualitative results: CIA-Net’s submitted entry surpasses all other methods in the 2018 MICCAI Multi-Organ Nuclei Segmentation Challenge.The figure displays the top 20 of 36 methods, and qualitative examples show less over- or under-segmentation in clustered and attached nuclei.
4 Conclusion
CIA-Net is proposed for instance-level nuclei segmentation, combining contour-aware information aggregation with a smooth truncated loss to address clusters and labeling noise. The conclusion positions the method as adaptable beyond nuclei segmentation to other medical image tasks.
- 4 Conclusion: CIA-Net is proposed to tackle prolific nuclei clusters and inevitable labeling noise in pathological images.The method is presented as a solution for instance-level nuclei segmentation.
- 4 Conclusion: The smooth truncated loss is part of CIA-Net’s approach to robust nuclei instance segmentation.The conclusion links the loss directly to handling clusters and labeling noise.
- 4 Conclusion: CIA-Net can be adapted to medical image segmentation tasks such as histology gland segmentation.The conclusion states that this adaptation can boost performance.