TL;DR
This paper introduces an automated method for accurately segmenting cell nuclei in H&E-stained histopathological images, a crucial step for disease diagnosis. The approach combines color deconvolution, morphological operations, and a novel marker-controlled watershed transform with a refinement step to effectively handle touching and overlapping nuclei, outperforming existing methods in accuracy and separation.
What Problem Does This Paper Address?
Pathologists manually examine histopathological slides, a process that is time-consuming, subjective, and prone to inter-observer variability. This is particularly challenging with tumor tissues due to their inherent heterogeneity and the complex, often touching or overlapping, nature of cell nuclei. Automated analysis of these images holds the promise of improving efficiency and accuracy in disease diagnosis, but it hinges on reliable cell nuclei segmentation. Existing automated methods struggle with variations in staining, uneven backgrounds, and the critical challenge of separating individual nuclei that are physically touching or overlapping.
Methods
The proposed method is a multi-stage pipeline designed to tackle the challenges of H&E-stained images:
- Color Deconvolution: The initial step uses a color deconvolution algorithm to separate the H&E-stained image into its constituent hematoxylin and eosin channels. This is crucial because hematoxylin stains the cell nuclei, making it the primary channel of interest.
- Morphological Operations and Thresholding: The hematoxylin channel undergoes morphological operations (opening by reconstruction and closing by reconstruction) to smooth nuclei boundaries and reduce noise. This is followed by an automated thresholding technique (Otsu's method) to create an initial binary image that roughly outlines the nuclei.
- Marker-Controlled Watershed Transform: This standard segmentation technique is employed to partition the image. However, to guide the watershed transform and prevent oversegmentation, "markers" are used. Internal markers (representing cell nuclei) are derived from the distance transform and H-maxima transform of the smoothed binary image. External markers (representing background) are derived from the morphological skeleton of the background.
- Segmentation Refinement (Novelty): This is the core of the paper's contribution. To address touching and overlapping nuclei that the initial watershed transform might not fully separate, a refinement step is introduced.
- Gradient-Weighted Distance Transform: This transform is applied to the segmented regions to highlight boundaries between touching nuclei.
- K-Means Clustering and Minimum Spanning Tree: Regional maxima from the gradient-weighted distance transform are clustered using k-means. The points within each cluster are then connected by a minimum spanning tree to form refined internal markers.
- Objective Function Minimization: The number of nuclei within a connected region is estimated by minimizing an objective function that considers the elliptical approximation of segmented regions. This objective function helps determine the optimal number of clusters (and thus, nuclei) to split a complex region into.
Datasets
The proposed method was evaluated on 10 H&E-stained histopathological images of lung cancer tissues, randomly selected from The Cancer Genome Atlas (TCGA) dataset. These images were obtained at a 40x magnification with a resolution of 0.25 µm/pixel. For ground truth, a subset of these images was manually segmented by an experienced pathologist.
Key Insights and Ideas
- Hematoxylin Channel is Key: The paper emphasizes that the hematoxylin channel, obtained through color deconvolution, is the most informative for nuclei segmentation due to its specific staining of nuclei.
- Marker-Controlled Watershed as a Foundation: While not entirely novel, using marker-controlled watershed is a robust starting point for segmentation, but it requires careful marker selection.
- The Power of Refinement: The most significant insight is the development of a sophisticated refinement step that specifically targets the challenging problem of separating touching and overlapping nuclei. This involves a clever combination of gradient information, clustering, and an objective function to estimate the true number of nuclei.
- Elliptical Approximation as a Heuristic: The assumption that cell nuclei tend to be somewhat elliptical is leveraged in the objective function to guide the separation process, which is a pragmatic approach given the inherent variability.
- Robustness to Parameter Variations: The authors highlight that some parameters, particularly those related to structuring element size, show robustness, which is crucial for a practical automated system.
Key Results
- Superior Segmentation Accuracy: The proposed method achieved a Dice Similarity Coefficient (DSC) of 0.880, significantly outperforming existing methods (Ref. 5: 0.815, Ref. 29: 0.794, Ref. 56: 0.755).
- Improved Sensitivity and PPV: The method also demonstrated higher Sensitivity (SN) of 0.931 and Positive Predictive Value (PPV) of 0.985 compared to the other methods.
- Enhanced Nuclei Separation: Quantitative evaluation of nuclei separation showed superior performance in reducing undersplit, oversplit, and encroachment errors (Undersplit: 0.076, Oversplit: 0.022, Encroachment error: 0.098).
- Visual Evidence: Figure 7 and 8 provide compelling visual evidence of the method's ability to accurately segment nuclei and effectively separate touching ones, often outperforming the compared methods.
- Computational Cost: While the proposed method is not the fastest, its average consumption time (4.05s per image) is competitive and justified by its improved accuracy.
Limitations and Future Work
- Elliptical Shape Assumption: The refinement step relies on an elliptical approximation of nuclei. While generally effective, it might struggle with nuclei that deviate significantly from this shape, though the authors note this is uncommon.
- Computational Time: The method's computational time is higher than some simpler approaches, though this is a trade-off for increased accuracy.
- Future Work: The authors suggest several promising avenues:
- Incorporating nucleus size into the objective function to further improve separation for non-elliptical shapes.
- Leveraging the segmentation results for cell classification using traditional methods or deep learning approaches.
- Applying the algorithm to cell tracking in time-lapse images for studying cell cycle progression and drug effects.
- Adapting the method for lower-resolution images (e.g., 20x magnification) by adjusting structuring element sizes.
Summary updated