Fig 1.
Orthogonal image slices of a 3D cell culture recording.
The images are grouped in two triple-sets, each displaying xy, zy and zx views of the same data set. The left side shows a slice at the z-depth 40, and the right side shows a slice at the z-depth 70. The images show a varying quality dependent on the region and a reduced sharpness in the z-axis as a result of the PSF being more spread out in this direction.
Fig 2.
Concept of the data synthesis pipeline.
The pipeline consists of four steps: dataset preparation, prototype generation, imaging simulation, and optimization. In favor of a clear visualization, 3D images are shown as 2D slices.
Fig 3.
2D slice of a real, naive and optimized image (top) and corresponding enlarged orthogonal views (bottom).
While the shape of the cell culture roughly matches between real and synthetic data, the positions of the nuclei are different due to the generation process of the synthetic data.
Fig 4.
Normalized Wasserstein metric between intensity distributions of real and naive (orange bars, Naive) and real and optimized data (blue bars, Optimized).
The measure is calculated for three image regions: background outside and inside the spheroid, and foreground. The images are divided into an upper, middle and lower part. A value of one indicates a perfect result, while a value of zero indicates a bad result.
Fig 5.
Difference in q95 edge quality between real and naive (orange bars, Naive) and real and optimized data (blue bars, Optimized).
Four corresponding images are compared, and the results for the normalized z-indices are grouped into the upper, middle, and lower region.
Fig 6.
Performance of segmentation models on real data.
The models are trained on naive and optimized image data. The left figure shows the detection accuracy (DET-Score), while the right figure shows the individual components of the metric: the number of false positive detections (FP), the number of false negative detections (FN), and the number of splitting operations (SplitOps).
Fig 7.
Zoomed 2D slices of a real image (upper left, Input), the corresponding ground truth (upper, middle, center point annotations) and segmentation results.
As segmentation methods, the classical algorithm TWANG and deep learning algorithms trained with naive, optimized, and the combination are used as indicated. The center point annotations are enlarged by five repeated morphological dilation operations with a diamond shaped structuring element. The size of the center points therefore corresponds to their distance to the shown slice.