vivoBodySeg: Machine learning-based analysis of C. elegans immobilized in vivoChip for automated developmental toxicity testing

Abstract

Developmental toxicity (DevTox) tests evaluate the adverse effects of chemical exposures on an organism's development. While large animal tests are currently heavily relied on, the development of new approach methodologies (NAMs) is encouraging industries and regulatory agencies to evaluate these novel assays. Several practical advantages have made C. elegansa useful model for rapid toxicity testing and studying developmental biology. Although the potential to study DevTox is promising, current low-resolution and labor-intensive methodologies prohibit the use of C. elegans for sub-lethal DevTox studies at high throughputs. With the recent availability of a large-scale microfluidic device, vivoChip, we can now rapidly collect 3D high-resolution images of ~ 1,000 C. elegans from 24 different populations. In this paper, we demonstrate DevTox studies using a 2.5D U-Net architecture (vivoBodySeg) that can precisely segment C. elegans in images obtained from vivoChip devices, achieving an average Dice score of 97.80. The fully automated platform can analyze 36 GB data from each device to phenotype multiple body parameters within 35 min on a desktop PC at speeds ~ 140x faster than the manual analysis. Highly reproducible DevTox parameters (4-8% CV) and additional autofluorescence-based phenotypes allow us to assess the toxicity of chemicals with high statistical power.

Keywords: C. elegans; U-Net; developmental toxicity; few-shot learning; high-throughput screening; microfluidics.

Figure 1. High-resolution C. elegans body images are acquired using a vivoChip-24x platform.

(a) Schematic of vivoChip-24x technology to immobilize C. elegans and capture their high-resolution images from 24 different populations and 40 worms per population. (b) Schematic of 1 well out of the 24 wells. The top of the device has a well. Underneath each well, there are 40 parallel immobilization channels. (c) Brightfield image of 40 adult C. elegans immobilized inside microfluidic channels within vivoChip-24x-3L device. Scale bar is 1 mm. (d) Graphic to demonstrate a z-stack of 10 images collected for each FOV (8 channels per FOV with 10x, 0.4 NA objective). For each FOV, we collect 50 images over 5 time points and 10 z-stacks (z-step size of 6 μm) per time point (1 s time interval). (e) We crop each channel for the first time point. (f) We identify the best focal plane for analysis and use 3 slices, including the image below and above the focal plane to train the network. (g) The location of the worm within the channel is used to determine the channel height based on its position with respect to the fiduciary cross mark present within each well of the vivoChip-24x.

Figure 2. The U-Net architecture of vivoBodySeg for automated analysis of C. elegans body.

(a) Schematic of our 4-layer U-Net. (b) Overview of our basic convolutional layer that defines our computation in our encoder and decoder. (c) Overview of vision transformer (viT).

Figure 3. Comparison of Dice scores for body segmentation by multiple scorers and by different vivoBodySeg models.

(a) Dice scores for each possible pair of five individual scorers for a subset of 20 channels. (b) The Dice score for all 301 test samples for three vivoBodySeg models arranged with worm numbers representing low to high Dice scores. The grey area represents the mean ± 2×standard deviation (μ ± 2σ) values for the Dice score from the segmentations of 5 individual scorers. Using the Wilcoxon signed rank sum test, the mean Dice score of vivoBodySeg-2.5D, Att was significantly higher than that of vivoBodySeg-2D and vivoBodySeg −2D, Att (p-value <0.001).

Figure 4. Images of C. elegans with developmental defects captured using two vivoChip-24x device designs.

(a) Image of 40 trapping channels from the 3-layer chip (vivoChip-24x-3L) with immobilized worms from a control population boxed in green (left panel) and a population that was treated with 4 μM CH₃Hg in 0.2% DMSO boxed in red (right panel). (b) Image of 40 trapping channels from a 4-layer chip (vivoChip-24x-4L) with immobilized worms from a control population boxed in green (left panel) and a population that was treated with 8 μM CH₃Hg in 0.2% DMSO (right panel). Scale bar = 1 mm. (c, d) High-magnification images of a single control worm (green box) and a CH₃Hg-treated worm (red box). Scale bar = 200 μm.

Figure 5. Improved worm detection using the few-shot learning of vivoBodySeg model. Images of a young L4

(a) and an adult stage (b) C. elegans, imaged in the vivoChip-24x-4L device. (c, d) Results from the baseline vivoBodySeg-2.5D, Att model with zero-shot learning. The model wrongly classified the channel as having a partial worm and thus discarded it for further analysis. (e, f) Results from the vivoBodySeg-2.5D, Att model with few-shot learning using 512 images. The model correctly predicted the worm mask and classified the channel images. Raw image, predicted mask, binary mask, and skeleton are compared. The scale bar is 200 μm.

Figure 6. Automated body parameter and autofluorescence analysis using the vivoBodySeg model.

(a) Map showing exposure of 24 C. elegans populations with CH₃Hg (0.5 – 9.0 μM) and 0.2% DMSO. The experiment was repeated with 5 vivoChip-24–4L devices. The data from all 5 experiments were used to quantify average body length (b), body area (c), and body volume (d). The data are presented as mean ± SEM (n = 5 repeats). The solid line represents the Hill-function fit to the experimental data. The dotted line represents the EC₁₀ for each parameter. The coefficient of variation (CV) is indicated for the control wells. (e) Example images of worms from 0.5 μM and 9.0 μM CH₃Hg. Brightfield (BF) and autofluorescence (AF) images obtained using the GFP filter are shown. The scale bar is 200 μm. (f) Average autofluorescence intensity per unit body length for CH₃Hg and DMSO control populations. The data are presented as mean ± SEM (n ≥ 4 repeats). The dotted line represents the EC₁₀ value. The solid line represents the LOAEL value of 1.0 μM for length, area, and volume (b – d) and 2.5 μM for Avg int/length (f).