Mathematica 11

We compared the nuclear segmentation efficacy of the pre-trained, deep-learning-based segmentation network with two conventional open-access software packages based on 3D watershed, which were previously published. We used 3D confocal stacks from untreated co-cultured KP4 spheroids with a diameter of 300 μm and processed them initially using Fiji. Pre-processing included the cropping of images, correcting the background, and exporting as tiff files. Conventional method 1 included an image analysis pipeline based on Mathematica (Mathematica 11.1, Wolfram Research Inc.) [32 (link)]. The median filter range was set to 3 pixels, the local threshold range was set to 10 pixels, and the hole filling range was set to 1 pixel. For seed detection, Laplace of Gaussian (LoG) was chosen with a seed range between 9 and 25 pixels. The other parameters were used by default. The results of the initial and final segmentation and the detected seed positions for different planes in xy and zy were displayed in real-time to increase the accuracy of the segmentation results. The post-segmentation data were exported as different 3D stacks in tiff format and as an XLSX file with quantitative results. For conventional method 2, we utilized OpenSegSPIM, an automated quantitative analysis tool for 3D microscopy data [33 (link)]. The parameter settings for the median filter and noise removal size were set automatically after the average nucleus size was measured using the built-in module. For the detection step, we selected shape-based detection with a sensitivity of 1 and started the segmentation process using shape-based segmentation. The segmentation masks were exported as different 3D stacks in tiff format and as an Excel sheet with quantitative results. The CNN segmentation was performed by a deep-learning model presented by Scherr et al. [34 (link)]. The structure of the model and the data processing routines were adapted to allow for a direct 3D processing of the data. For training and inference, the images were sliced into patches of size 32 × 128 × 128 px³ (z, y, x). Training was performed with the ranger optimizer, an initial learning rate of 1 × 10^− 3 and a batch size of three. Other training parameters were consistent with the ones used in [34 (link)]. The development of a substantial deep-learning training dataset for 3D data presents significant challenges. The extra dimension, relative to 2D data, substantially lengthens labeling time. Additionally, distinguishing nuclei in high-density areas is complex and labor-intensive, owing to reliance on 2D visualizations. Therefore, four synthetic 3D spheroid images, generated as described in [35 (link)] were used for model training. Quantitative analysis of the segmentation performance was conducted with the segmentation and detection measures used in the cell tracking challenge [36 (link)]. As ground truth, an image patch of size 32 × 128 × 128 px³ was manually annotated by a biological expert. The image patch was extracted from the 5 images.

To investigate the parallels between morphological alterations in mice and their potential manifestations in humans, we conducted a detailed transformation of polygon data representing human craniofacial structures. The original image was extracted from full-body MRI scans (from the BodyParts3D dataset, a resource developed by the Database Center for Life Science in Tokyo, Japan) and transformed as specified below. Our focus was on the human skull, which was dissected into 53 high-resolution segments encompassing an array of teeth, bones, and ligaments.
For the transformation process, we employed the capabilities of Mathematica 11.0, (developed by Wolfram Research in Illinois, USA) using custom-written code (available at https://zenodo.org/records/10363659). Our transformation technique involved a non-linear 3D transformation algorithm conceptualized as a three-dimensional ‘magnifying glass.’ This method was adapted from previously published code⁷⁹ and featured a radius of 3 cm centered around the nasal cavity.
We calibrated the magnification effect to mirror the expansion observed in the nasal cavity of mice. This allowed us to extrapolate and hypothesize the potential craniofacial changes in humans corresponding to those noted in mouse models. Our algorithm was designed to allow for selective magnification of specific anatomical regions within the skull. This approach ensured that while we magnified certain areas for detailed study, the rest of craniofacial anatomy remained undistorted and true to its original proportions.