Cell segmentation and representation with shape priors.

Area of Science:

• Computational biology
• Bioimage analysis
• Machine learning for biology

Background:

• Cell segmentation is crucial in computational biology, with deep learning, specifically convolutional neural networks, achieving state-of-the-art results.
• Shape-constrained segmentation models are emerging as powerful alternatives to traditional methods for instance segmentation.
• The choice of shape representation is critical for the efficacy of shape-constrained segmentation.

Purpose of the Study:

• To introduce and evaluate two novel representation-based deep learning segmentation methods.
• To quantitatively compare existing shape descriptors for cell segmentation.
• To assess the performance of new methods against established deep learning baselines.

Main Methods:

• Development of two deep learning segmentation networks utilizing Fourier coefficients and statistical shape models.
• Quantitative comparison of prominent shape descriptors from existing literature.
• Implementation of shape-constrained deep learning for cell instance segmentation.

Main Results:

• The proposed methods demonstrate competitive performance compared to widely used deep learning algorithms.
• Effectiveness is particularly noted when using a low number of parameters for the shape model.
• The methods show promise for segmenting cells with irregular morphologies.

Conclusions:

• The introduced Fourier coefficient and statistical shape model-based networks provide viable alternatives for cell segmentation.
• These shape-constrained deep learning approaches are especially beneficial for complex cellular shapes and efficient model parameterization.
• The study highlights the importance of shape representation in advancing deep learning-based cell segmentation.