Improving and evaluating deep learning models of cellular organization

Cells contain dozens of major organelles and thousands of other structures, many of which vary extensively in their number, size, shape and spatial distribution. This complexity and variation dramatically complicates the use of both traditional and deep learning methods to build accurate models of cell organization. Most cellular organelles are distinct objects with defined boundaries that do not overlap, while the pixel resolution of most imaging methods is n sufficient to resolve these boundaries. Thus while cell organization is conceptually object-based, most current methods are pixel-based. Using extensive image collections in which particular organelles were fluorescently labeled, deep learning methods can be used to build conditional autoencoder models for particular organelles. A major advance occurred with the use of a U-net approach to make multiple models all conditional upon a common reference, unlabeled image, allowing the relationships between different organelles to be at least partially inferred.

We have developed improved Generative Adversarial Networks-based approaches for learning these models and have also developed novel criteria for evaluating how well synthetic cell images reflect the properties of real images. The first set of criteria measure how well models preserve the expected property that organelles do not overlap. We also developed a modified loss function that allows retraining of the models to minimize that overlap. The second set of criteria uses object-based modeling to compare object shape and spatial distribution between synthetic and real images. Our work provides the first demonstration that, at least for some organelles, deep learning models can capture object-level properties of cell images.

http://murphylab.cbd.cmu.edu/Software/2022_insilico.

Supplementary data are available at Bioinformatics online.