[Paper Review] Constrained Multi-shape Evolution for Overlapping Cytoplasm Segmentation

Youyi Song, Lei Zhu+5 more|arXiv (Cornell University)|Apr 8, 2020

AI in cancer detection39 references4 citations

TL;DR

This paper proposes a constrained multi-shape evolution method for segmenting overlapping cytoplasm in cervical smear images by jointly evolving multiple cytoplasm shapes using both local (cytoplasm-level) and global (clump-level) shape priors. It models an infinitely large shape hypothesis set from statistical shape data, enforces shape constraints during evolution, and achieves superior segmentation accuracy over state-of-the-art methods by compensating for intensity deficiency in overlapping regions.

ABSTRACT

Segmenting overlapping cytoplasm of cells in cervical smear images is a clinically essential task, for quantitatively measuring cell-level features in order to diagnose cervical cancer. This task, however, remains rather challenging, mainly due to the deficiency of intensity (or color) information in the overlapping region. Although shape prior-based models that compensate intensity deficiency by introducing prior shape information (shape priors) about cytoplasm are firmly established, they often yield visually implausible results, mainly because they model shape priors only by limited shape hypotheses about cytoplasm, exploit cytoplasm-level shape priors alone, and impose no shape constraint on the resulting shape of the cytoplasm. In this paper, we present a novel and effective shape prior-based approach, called constrained multi-shape evolution, that segments all overlapping cytoplasms in the clump simultaneously by jointly evolving each cytoplasm's shape guided by the modeled shape priors. We model local shape priors (cytoplasm--level) by an infinitely large shape hypothesis set which contains all possible shapes of the cytoplasm. In the shape evolution, we compensate intensity deficiency for the segmentation by introducing not only the modeled local shape priors but also global shape priors (clump--level) modeled by considering mutual shape constraints of cytoplasms in the clump. We also constrain the resulting shape in each evolution to be in the built shape hypothesis set, for further reducing implausible segmentation results. We evaluated the proposed method in two typical cervical smear datasets, and the extensive experimental results show that the proposed method is effective to segment overlapping cytoplasm, consistently outperforming the state-of-the-art methods.

Motivation & Objective

• To address the challenge of overlapping cytoplasm segmentation in cervical smear images, where intensity information is often insufficient due to overlapping regions.
• To overcome limitations of existing shape prior-based methods that use limited shape hypotheses and lack constraints on resulting shapes.
• To improve segmentation accuracy by jointly evolving multiple cytoplasm shapes using both local and global shape priors.
• To ensure physically plausible segmentation results by constraining evolved shapes within a learned, comprehensive shape hypothesis set.

Proposed method

• The method models local shape priors using an infinitely large shape hypothesis set derived from statistical shape analysis of cytoplasm instances.
• It introduces a learning algorithm to compute shape example importance, enabling the hypothesis set to represent all possible cytoplasm shapes.
• Global shape priors are incorporated by modeling mutual shape constraints among cytoplasms within a cell clump, improving consistency in overlapping regions.
• During evolution, both local and global shape priors are used to guide shape evolution, compensating for intensity deficiency.
• The resulting shape at each evolution step is constrained to remain within the learned shape hypothesis set to prevent implausible shapes.
• A convergence analysis is provided, proving that the importance learning algorithm converges to the optimal weights regardless of input order.

Experimental results

Research questions

• RQ1 Can a comprehensive shape hypothesis set that includes all possible cytoplasm shapes improve segmentation accuracy in overlapping regions?
• RQ2 How does incorporating global shape priors—modeling mutual constraints among cytoplasms in a clump—affect segmentation quality?
• RQ3 Can constraining evolved shapes to lie within a learned shape hypothesis set reduce implausible segmentation results?
• RQ4 Does the proposed method outperform existing shape prior-based approaches in terms of accuracy and visual plausibility on overlapping cytoplasm segmentation?
• RQ5 Is the importance learning algorithm for shape statistics robust and convergent under varying input orders?

Key findings

• The proposed method consistently outperforms state-of-the-art methods on two typical cervical smear image datasets in terms of segmentation accuracy.
• The use of both local and global shape priors significantly improves segmentation in intensity-deficient overlapping regions.
• Constraining evolved shapes within the learned shape hypothesis set effectively reduces visually implausible segmentation results.
• The importance learning algorithm is proven to converge to optimal weights, with convergence independent of input order.
• The method achieves superior performance by jointly evolving multiple cytoplasm shapes, preserving structural consistency across the clump.