DARC: Distribution-Aware Re-Coloring Model for Generalizable Nucleus Segmentation

Shengcong Chen, Changxing Ding, Dacheng Tao, Hao Chen

TL;DR

This paper introduces DARC, a novel framework for nucleus segmentation that tackles domain shift challenges in pathological images. DARC addresses variations in image acquisition and, crucially, the impact of foreground-background ratio on feature statistics, leading to more robust and generalizable models.

What Problem Does This Paper Address?

Accurate nucleus segmentation is vital for analyzing pathological images. However, a significant hurdle is the "domain gap" – variations in how images appear due to different scanners, staining techniques, and tissue types. This variability makes models trained on one dataset perform poorly on another. While domain adaptation methods exist, they often require target domain data, which is not always available in real-world scenarios. This paper argues that domain gaps are exacerbated by the varying ratio of foreground (nuclei) to background pixels, which can skew feature statistics, particularly for normalization layers. Existing domain generalization (DG) techniques, often designed for classification, may not fully address these specific challenges in dense prediction tasks like segmentation.

Methods

The proposed DARC model builds upon a U-Net architecture, a standard for segmentation tasks. The core innovations lie in two components:

  • Re-Coloring (RC) Module: This module aims to mitigate the visual differences caused by varying image acquisition conditions. It first converts the RGB image to grayscale and then uses a simple network to generate an initial re-colored image. To preserve fine-grained texture information lost during de-colorization, it employs a "Sort-Matching" algorithm to re-integrate semantic information from the original image.
  • Distribution-Aware Instance Normalization (DAIN): Recognizing that standard Instance Normalization (IN) can be sensitive to foreground-background ratio variations, DAIN is introduced. It re-estimates feature statistics by explicitly considering this ratio. The model first predicts the foreground-background ratio ($\rho$) using a Ratio Prediction Head (RPH). This predicted ratio is then used to adjust the feature statistics, making the normalization more robust. The RPH is trained with a combined binary cross-entropy and mean squared error loss.

DARC integrates these components into an end-to-end trainable model. Different configurations are explored, including DARCall (all normalization layers replaced by DAIN) and DARCenc (DAIN in the encoder, standard Batch Normalization in the decoder).

Datasets

The paper evaluates DARC on four publicly available datasets:

  • H&E stained: CoNSeP, CPM17
  • IHC stained: DeepLIIF, BC-DeepLIIF

These datasets represent diverse image acquisition conditions and staining protocols, making them suitable for evaluating domain generalization capabilities.

Key Insights and Ideas

  • Foreground-Background Ratio is Crucial: The paper highlights that the ratio of foreground to background pixels is a significant, often overlooked, factor contributing to domain shift in segmentation tasks. This ratio directly impacts feature map statistics and the effectiveness of normalization layers.
  • Instance Normalization's Sensitivity: Standard Instance Normalization, while better than Batch Normalization for domain generalization, is still susceptible to variations in foreground-background ratios, necessitating a more adaptive approach like DAIN.
  • Re-coloring Preserves Structure: The proposed Re-Coloring module effectively reduces visual domain gaps while crucially preserving fine-grained structural information, which is essential for accurate segmentation.
  • Encoder-Decoder Architecture Matters: The DARCenc configuration, which uses DAIN in the encoder and standard BN in the decoder, achieves the best performance, suggesting that robust feature extraction in the encoder is paramount, and a well-trained decoder can leverage these features effectively.

Key Results

DARCenc demonstrated superior performance across all tested datasets, achieving the best average AJI and Dice scores. Specifically, DARCenc improved the baseline model's average AJI and Dice scores by 4.81% and 7.04%, respectively. The paper also shows that the Re-Coloring module alone improved average AJI and Dice scores by 1.41% and 2.59%, while DAIN improved them by 1.13% and 4.08%. The combination of RC and DAIN (DARCall and DARCenc) yielded the most significant gains. Qualitative results (Figure 3) visually confirm DARCenc's ability to produce more accurate segmentations compared to other methods.

Limitations and Future Work

  • Computational Complexity: While DARC achieves better performance, Table 3 indicates a slight increase in parameter count and inference time compared to the baseline, though still competitive.
  • Future Work: The paper implicitly suggests that further exploration of adaptive normalization techniques and their integration into different network architectures could lead to even more robust domain generalization for medical image analysis. The success of DARCenc also points towards optimizing the balance between adaptive normalization in the encoder and standard normalization in the decoder.

Summary updated

Abstract

Nucleus segmentation is usually the first step in pathological image analysis tasks. Generalizable nucleus segmentation refers to the problem of training a segmentation model that is robust to domain gaps between the source and target domains. The domain gaps are usually believed to be caused by the varied image acquisition conditions, e.g., different scanners, tissues, or staining protocols. In this paper, we argue that domain gaps can also be caused by different foreground (nucleus)-background ratios, as this ratio significantly affects feature statistics that are critical to normalization layers. We propose a Distribution-Aware Re-Coloring (DARC) model that handles the above challenges from two perspectives. First, we introduce a re-coloring method that relieves dramatic image color variations between different domains. Second, we propose a new instance normalization method that is robust to the variation in foreground-background ratios. We evaluate the proposed methods on two H$\&$E stained image datasets, named CoNSeP and CPM17, and two IHC stained image datasets, called DeepLIIF and BC-DeepLIIF. Extensive experimental results justify the effectiveness of our proposed DARC model. Codes are available at \url{https://github.com/csccsccsccsc/DARC
BibTeX
@inproceedings{chen2023darc,
  author = {Shengcong Chen and Changxing Ding and Dacheng Tao and Hao Chen},
  title = {{DARC: Distribution-Aware Re-Coloring Model for Generalizable Nucleus Segmentation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.1007/978-3-031-43987-2\_57},
}
Figure