Page 182 - IJB-8-1

Analyzing Cell-Scaffold Interaction through Unsupervised 3D Nuclei Segmentation
individual cells or ellipsoids-like nuclei, that is, cell/nuclei discriminator are trained to compete against each other.
segmentation . As such, several rule-based methods have The former learns to create images that look real, and
[4]
been developed, which commonly apply visual features the latter learns to distinguish real or fake images. Hou
[18]
to discriminate regions-of-interest on microscopy images. et al. proposed a GAN-based method to refine synthetic
For example, a global threshold is utilized to determine histopathology nuclei images. The pairs of the refined
if one individual image pixel is an object or not. This synthetic nuclei images and the corresponding masks
approach is easy to implement but fails to tackle noisy were used to train a task-specific segmentation model,
images or de-convolute clustered nuclei [5,6] . Meanwhile, which is too complicated for practical applications. In
how to extract and select features is always arguable . most cases, translating images cross domains requires
[7]
Marker-based watershed segmentation is capable of datasets with paired information, which may not be
separating cells, but the relevant marker detection highly available in real-world applications. To enforce stronger
[16]
depends on parameter selection [8,9] . Standard edge- cross-domain connection, Dunn et al. proposed a cycle-
detection algorithms can only discriminate the lateral consistent adversarial network (CycleGAN) to translate
boundaries under low cell density . Despite the great the synthetic semantic label maps from mask style to
[10]
interest and expectation from the research community fluorescence microscopy image style and use the pairs
on these rule-based methods, they are not scalable when for training a segmentation model. Unfortunately, this
analyzing a large number of cells , or clustered cells CycleGAN cannot achieve one-to-one nuclei mapping
[11]
with weak boundary gradients. To meet the growing between the images and corresponding masks.
demand from biological image analysis, some stand-alone Content and style are the two most inherent attributes
commercial software has been developed to segment to characterize nuclei visually. Recent achievement shows
nuclei such as CellProfiler and Squassh [13] in ImageJ. In that disentangling content and style representations in
[12]
general, they are convenient to use but lack of flexibility image-to-image translation can significantly improve
[19]
when dealing with 3D segmentation applications. GAN’s performance. Yang et al. applied this technology
Over the past decade, machine learning (ML) to attain cross-modality domain adaptation between
has been introduced to segment CLSM images and computed tomography and magnetic resonance imaging
has demonstrated state-of-the-art performance under images using a shared content space. Inspired by this idea,
varied image resolution, signal-to-noise ratio, contrast, we apply this disentanglement technology to align the
and background [14,15] . Supervised ML-based nuclei domains’ content representation under a novel end-to-end
segmentation methods consist of three key elements: model, called Aligned Disentangling GAN (AD-GAN). As
derived staining patterns, extracted features of nuclei shown in Figure 1, two style representations are involved
morphology, and annotated training data. The current in this AD-GAN model with different structure rendering,
bottleneck of applying such methods lies in the expensive that is, image and mask. This AD-GAN can extract content
and tedious process of appropriate annotating thousands and style representation from CLSM image separately. As
of nuclei with irregular or deformed shapes as training a result, CLSM images can be disentangled into content
data. Particularly, the poor Z axial resolution of CLSM representation (underlying nuclei spatial structure) and
images causes difficulties and defects when delineating image style (rendering of the structure). As shown on the
the top and bottom boundaries of nuclei . right side of Figure 1, the same content representation can
[16]
To deal with this dilemma, unsupervised ML
methods have been introduced for nuclei segmentation.
Unsupervised methods can learn relatively consistent
pattern from large scale of data without expensive
annotation and personal bias. They can make use of
prior knowledge from human beings at abstract level,
which may be closer to the essence of learning. Thus,
unsupervised methods with appropriate prior knowledge
can obviously minimize the influence from the poor Z-axis
resolution, in comparison with supervised ML methods
using concrete annotation. Liu et al. took advantages
[17]
of unsupervised domain adaptation (UDA) to transfer
knowledge of nuclei segmentation from fluorescence
microscopy images to histopathology images. However,
this UDA requires labeled fluorescence microscopy
images. Generative adversarial network (GAN) has also Figure 1. Principle of nuclei segmentation using aligned
been introduced to label images, where a generator and a disentangled generative adversarial network.

168 International Journal of Bioprinting (2022)–Volume 8, Issue 1