An energy conduction model for cell image segmentation

Ma Jingfeng, Jiajun Bu, Kai Hou, Shanglian Bao, Chun ChenChinese 2011 · DOI

Abstract Cell image segmentation is an essential step in cytopathological analysis. Although their execution speed is fast, the results of cell image segmentation by conventional pixel-based, edge-based and continuity-based methods are often coarse. Fine structures in a cell image can be obtained with a method that quickly adjusts the threshold levels. However, the processing time of such a method is usually long and the final results may be sensitive to intensity differences and other factors. In this article, a new energy model is proposed that synthesizes a differential equation from the conventional and level set methods, and utilizes the nonuniformity property of cell images (e.g. cytoplasms are more uneven than the background). The feasibility and robustness of the proposed model was demonstrated by processing relatively complicated background images of both simulated and real cell images.

Introduction

In cytopathology—the study of disease at a cellular level—segmenting a cell into its constituent parts, namely the nucleus and cytoplasm, is the critical first step for diagnosis. However, cell images are notoriously difficult to process. Conventional pixel-based or edge-based methods are fast but produce "coarse" results that miss fine structures. More advanced "active contour" or "level set" models, which treat the segmentation boundary like a flexible rubber band that shrinks to fit an object, are more precise but often stumble when the "signal" is weak. Specifically, if the intensity difference between the cell and the background is minor, the mathematical "energy" driving the contour lacks a clear gradient to follow, leading to failed segmentations.

Methodology

The proposed energy conduction model is implemented within a level set framework. This approach relies on a key biological observation: while a cell and its background might have similar brightness (intensity), they possess very different textures. The cytoplasm is "uneven" due to various organelles, meaning it has a much higher intensity gradient than the relatively smooth background.

Conceptual model and image processing: (a) Domain definitions for background, cytoplasm, and nucleus; (b) Original cell image; (c) Gradient-enhanced image highlighting cellular nonuniformity.

Technical Innovations

  1. Feature Fusion ($IG$): The researchers create a composite image ($IG$) by blending the original image ($I$) with its gradient ($G$). This amplifies the "energy" difference between the cell and the background, making the boundary mathematically distinct.
  2. Two-Stage Pipelining: Instead of solving the whole image at once, the energy function is split into two separate, sequential tasks. Stage 1 segments the entire cell (cytoplasm + nucleus) from the background using the composite $IG$ image. Stage 2 then segments the nucleus from the cytoplasm using the original intensity image $I$, as the nucleus is typically much denser and darker than the surrounding cytoplasm.
  3. Region-Scalable Fitting (RSF): By incorporating a Gaussian kernel, the model handles intensity inhomogeneity (uneven lighting), allowing it to adapt to local variations rather than relying on a global threshold.

Results and Discussion

The model was evaluated using simulated cell images to demonstrate robustness and real cervical carcinoma cell images acquired from laboratory settings.

Segmentation result showing the model's ability to delineate complex boundaries in high-texture environments.

Key Findings

  • Robustness to Low Contrast: In head-to-head tests, conventional level set methods failed to find the cell boundary due to insufficient intensity differences. The proposed model successfully segmented the same images by leveraging the gradient data.
  • Efficiency: Despite the mathematical complexity, the pipelined approach proved efficient. For images around 116×160 pixels, the total processing time was approximately 1.125 to 1.318 seconds on a standard workstation.
  • Sub-pixel Accuracy: Because the model uses a variational level set approach (solving differential equations) rather than just grouping pixels, it achieves segmentation results at a sub-pixel level, which is vital for quantitative shape analysis.

Conclusion

By treating the "noise" of the cytoplasm as a high-gradient signal, the energy conduction model can "see" cells that are otherwise invisible to intensity-only algorithms. While the model is faster than previous piecewise-smooth models, it still requires iterative calculations. Future work will explore parallel processing across GPUs to handle high-resolution pathology slides and further automate the initialization process to create a fully "hands-off" diagnostic pipeline.

Summary updated

Abstract

Cell image segmentation is an essential step in cytopathological analysis. Although their execution speed is fast, the results of cell image segmentation by conventional pixel-based, edge-based and continuity-based methods are often coarse. Fine structures in a cell image can be obtained with a method that quickly adjusts the threshold levels. However, the processing time of such a method is usually long and the final results may be sensitive to intensity differences and other factors. In this article, a new energy model is proposed that synthesizes a differential equation from the conventional and level set methods, and utilizes the nonuniformity property of cell images (e.g. cytoplasms are more uneven than the background). The feasibility and robustness of the proposed model was demonstrated by processing relatively complicated background images of both simulated and real cell images.
BibTeX
@inproceedings{jingfeng2011energy,
  author = {Ma Jingfeng and Jiajun Bu and Kai Hou and Shanglian Bao and Chun Chen},
  title = {{An energy conduction model for cell image segmentation}},
  booktitle = {Chinese},
  year = {2011},
  doi = {10.1007/s11434-011-4389-z},
}
Figure