Developmental growth, first set up.

Wild-type N2 C. elegans were grown at 20°C and were fed OP50 E. coli bacteria as described in Guisnet et al. [45]. Worms were timed by placing egg-laying adults on plates for 2 hours. These worms were moved to fresh plates through the experimental days as needed to keep them fed. Every 24 hours after that, worms were placed on 2% agarose pads in a 5 µL drop of 1M sodium azide to immobilize them. After 1 minute, the sodium azide was carefully absorbed with a Kimwipe. No coverslip was added to cover the worms as this caused them to be squished and deformed (particularly in width). Starting at day 3, focus was optimized for the head and tail as worms were too thick to be entirely in focus. Worms were imaged within 10 minutes of being immobilized. Grayscale images were taken with an Olympus XM10 CCD camera mounted on an Olympus MVX10 microscope using an MV PLAPO 2XC objective. TIF images were acquired with a 1376 x 1038 resolution and 10 ms exposure with the cellSens software (Evident).

Developmental growth, second set up.

C. elegans (N2 wild-type strain) were grown at 20°C on nematode growth medium (NGM) agar plates seeded with Escherichia coli OP50. Synchronized worms were obtained as previously described [46]. Animals were anesthetized in 0.2% (w/v) levamisole prepared in M9 buffer and mounted directly between a glass slide and a coverslip. Imaging was performed at room temperature using a Leica DMI6000B inverted microscope equipped with a 4 × /0.10 NA objective lens, a spinning-disk confocal head (Yokogawa CSU10), and an EM-CCD camera (Hamamatsu ImagEM). Image tiling was performed using Metamorph software, and subsequent stitching of the tiles was carried out in ImageJ.

Thresholding analysis.

For both set-ups, thresholding analysis was tested by using the Default and Otsu thresholding options of the Fiji software [33].

Automated multi-worm morphological feature extraction from images.

After image acquisition of developmental growth animals, we employed the Segment Anything Model (SAM) [19] to generate automatic mask predictions for each image in our dataset. The ViT-H variant of SAM was utilized, with the following custom parameters: points_per_side: 32, pred_iou_thresh: 0.95, stability_score_thresh: 0.95, min_mask_region_area: 1500. These parameters were empirically determined to optimize mask generation for worm-like objects. The SAM model was implemented using PyTorch [47] and Python 3.10.12 and executed on a CUDA-enabled GPU for efficient processing. For each mask predicted by SAM, we extracted a cutout from the original image. Masks touching image edges were excluded to avoid partial worms.

487 generated cutouts from the developmental growth, first set up acquisitions underwent manual classification by C. elegans experts. Each cutout was labeled as either “worm” or “not worm” using the LabelBox web tool with academic access [48]. The classified cutouts were randomly split into training (80%) and validation (20%) sets. We employed transfer learning using a Vision Transformer (ViT-H-14) [49] pre-trained on ImageNet [50] as our base model. The classifier was fine-tuned on our worm/notworm dataset. The “worm” classification included partial worms to improve future usability of the model, although these were later filtered out. The base model’s parameters were frozen, and only the final classification layer was retrained. Data augmentation techniques, including random resized crops and horizontal flips, were applied to the training set. The model was trained for 25 epochs using stochastic gradient descent (SGD) optimization with a learning rate of 0.001 and momentum of 0.9. A step learning rate scheduler was implemented, decaying the learning rate by a factor of 0.1 every 7 epochs. Cross-entropy loss was used as the optimization criterion. Training was conducted on a single GPU using PyTorch. The model achieving the highest validation accuracy (100%) was selected as the final model.

Pre-computed mask predictions from SAM were classified as “worm” or “not worm” by the fine-tuned classifier. To address instances of overlapping worm masks (that is, when there were multiple masks produced for the same worm representing subparts of the body), an adjacency matrix was constructed to identify overlapping regions. Connected components analysis was then performed to group overlapping masks. Within each group, only the mask with the largest area was retained for further analysis, ensuring each worm was represented only once. Touching worms were automatically resolved by SAM.

For each retained worm mask, we extracted the following morphological features: area (sum of pixels within the mask), perimeter (Euclidean number of pixels of the shape’s contour), 100-point interpolated medial axis (the longest path along the worm’s centerline identified through graph analysis), length (Euclidean pixel count of the medial axis), width (estimated at each point along the medial axis). For age-based identification in images with mixed-age worms, a size deviation threshold relative to all age-matched worms identified was applied per imaging day.

Noise degradation analysis.

The original 16-bit grayscale images were corrupted with additive Gaussian noise at six increasing intensity levels. Noise was sampled from a zero-mean Gaussian distribution with standard deviations (σ) of 655, 1,966, 3,276, 4,587, 6,553, and 13,107 pixel intensity units, corresponding to approximately 1%, 3%, 5%, 7%, 10%, and 20% of the full 16-bit range, respectively. Each set of degraded images was then processed using identical parameters as for the original dataset. Performance at each noise level was evaluated by comparison to the expert manual segmentation, using Pearson correlation coefficient and per-worm percent area agreement (TWARDIS area / manual area × 100). To distinguish between failure at different pipeline stages, false negative and false positive rates were computed separately for the segmentation step (whether SAM could produce a valid non-edge mask for the worm) and the classification step (whether our ViT classifier correctly classified the worm mask).

Runtime benchmarking.

21 microscopy images from the dataset were each processed at four spatial resolutions: 100%, 90%, 70%, and 50% of the original pixel dimensions, 1376x1038, 1238x934, 963x726 and 688x519 pixels, respectively, obtained by downscaling with area-based interpolation.

Runtime was measured for the three main pipeline steps: SAM2 automatic mask generation, ViT binary classification of each candidate mask, and (3) morphological metrics extraction for each detected worm. Metrics extraction was further decomposed into six sub-steps: connected component extraction, contour detection and perimeter measurement, medial axis skeletonization, skeleton endpoint and branchpoint cleaning, and width/length measurements along the pruned skeleton (worm spline).

GPU benchmarks were performed on an NVIDIA RTX 3090 (24 GB VRAM) using CUDA with bfloat16 autocast and TF32 precision. A warmup pass was run before timing. CPU benchmarks were performed on the same machine using 4, 8, and 16 threads, controlling parallelism through PyTorch threads option (torch.set_num_threads), OpenCV threads (cv2.setNumThreads), and environment variables (MKL_NUM_THREADS). The skeleton cleaning step was additionally parallelized using a thread pool (concurrent.futures.ThreadPoolExecutor). All timing measurements used time.perf_counter(). Classification was performed in-memory to avoid disk I/O artifacts in timing measurements.

Droplet swimming recording set up.

Wild-type N2 C. elegans were grown at 20°C and were fed OP50 E. coli bacteria as described in Guisnet et al. [45]. A large 10 cm diameter plexiglass dish that was first cleaned with ethanol and a Kimwipe. It was then grounded by bringing the surface in very close proximity to a metal cabinet. This removed static from the plexiglass helping small droplets to stick to the surface. 50 µL of filtered NGML was then pipetted onto the center of the dish. Young adult worms were selected from maintenance plates and left to roam off food on an empty agar plate for 2 minutes before being moved to the droplet with a platinum worm pick.

Uncompressed AVI videos were recorded for 30 seconds at 10 fps with a FLIR Blackfly (BFS-U3-51S5M-C) 5.0 MP camera and Spinnaker SDK software (Teledyne) at 2448 x 2048 resolution. The recording dish was held upside down with a custom made rig and recorded from below. An LED light ring of the same size as the plate was positioned centered with the plate, but further up (~ 8 centimeters), to minimize light refraction artifacts in the droplet and dish plastic.

Automated quantitative analysis of swimming behavior.

Worm segmentation was performed using the Segment Anything Model 2 (SAM2) [20]. The model was initialized using a pre-trained checkpoint (sam2_hiera_large.pt) and GPU acceleration was employed using CUDA, with TensorFloat-32 compatible hardware. A generic prompt frame was added to all videos and used to automatically guide segmentation on the worm object. Segmentation was propagated backwards from the added prompt frame through the video. To improve segmentation accuracy, a second pass of segmentation was performed on cropped regions of interest. The SAM2 model was applied again to these cropped frames, using a new generic prompt frame matching the crop size. Segmentation was again propagated backwards through the cropped frames. The resulting worm high-definition masks were used for further analysis.

SAM2 masks from the high-definition segmentation were cleaned using morphological operations to remove occasional small particles. Worm skeletons were extracted by mask skeletonization. A custom algorithm was implemented to correct self-touching skeletons, ensuring a single, continuous skeleton for each frame. Skeleton points were smoothed using spline interpolation to 100 points. In addition to those described in the morphological extraction section above, several other metrics were extracted from the skeletons. Gaussian-weighted curvature was calculated along the smoothed skeleton using a window size of 50 points and sigma of 10. Amplitude along the worm’s body was measured as the perpendicular distance from skeleton points to the worm’s centerline (defined as the straight line from head to tail). Worms were classified by shape. Classification was based on curvature and amplitude values (Table 4). Thresholds for classification were empirically determined. Worms were classified as turned on the z-plane, if their length was smaller by more than 1 standard deviation from the video mean. Wavelength was based on curvature peaks for S-shaped worms, set at 2 worm length for C-shapes, and none for straight or turned worms. Wave number was calculated as the ratio of worm length to wavelength. Spatial frequency analysis was performed using Fast Fourier Transform on the curvature data. Temporal frequencies were calculated from the curvature data. The power spectral density (PSD) of the normalized curvature time series was calculated using Welch’s method with a window size of 30 and an overlap of 25. A sliding window analysis was conducted to identify the dominant frequencies within each segment. Peaks in the PSD were detected, and the frequency corresponding to the peak with the maximum power was selected as the dominant frequency. If no peaks were found, the frequency with the maximum power was used. The dominant frequencies were interpolated across the entire time series to ensure a frequency value for each frame.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Shape classification.

https://doi.org/10.1371/journal.pcbi.1013441.t004

Crawling recording set up.

Wild-type N2 C. elegans were grown under standard conditions at 20°C and were fed OP50 E. coli bacteria as described in Guisnet et al. [45]. Young adult worms were selected from maintenance plates and left to roam off food on an empty agar plate for 2 minutes before being recorded. AVI videos were recorded for 1 minute with the same setup and settings as described for swimming.

Tierpsy analysis.

The version of Tierpsy available on April 23rd 2026 was installed via Docker and used through GUI by following the instructions presented on the GitHub page (https://github.com/Tierpsy/tierpsy-tracker) [51]. Thresholding parameters were selected to optimize for real worm detection, while minimizing false positives from background artifacts, by manually iterating through 5 randomly selected videos. These parameters were then used to carry out the analysis in batch on all 114 recordings. All Tierpsy defaults were maintained, except for the use_nn_filter option which was changed to False, since keeping this default option resulted in zero worm detection.

To quantify Tierspy analysis success, we used the timeseries_data from the featuresN output file. A frame was qualified having failed to be analyzed by Tierpsy if all features were NaN, regardless of how many or which features were provided.

Automated single worm high-definition tracking.

Worm tracking was performed in a similar way as described for swimming with a few differences for path tracking and addressing more complicated body postures. For the high-definition segmentation, two generic prompt frames with complicated shapes were added on cropped frames and the SAM2 model was reinitialized and applied to each of the cropped frames one by one. Segmentation was propagated through these short cropped sequences to generate the high-definition masks. The resulting cropped masks were then resized and mapped back to the original frame dimensions for further analysis.

Skeleton morphological features were extracted as described in the swimming section with additional extraction of head bends. Due to the low resolution of the worm in the large image space, head and tail differentiation was not possible by traditional method of brightness difference [37]. A temporal tracking algorithm was employed instead, utilizing a 5 frame sliding window to analyze movement patterns. Endpoint grouping was optimized using the Hungarian algorithm, with the more mobile group designated as the head based on cumulative displacement. Error correction was applied for sudden jumps by cross-referencing with recent history and worm movement. Then, head bending magnitude was quantified by calculating the angle between the head endpoint and the main body axis.

For path analysis, the segmentation masks from previous shape analysis were used to obtain the centroid position for every frame using center-of-mass computation. To reduce measurement noise inherent from changes in the worm’s shape, centroid trajectories were smoothed using Savitzky-Golay filtering, which preserves the underlying movement patterns while eliminating high-frequency noise. Worm movement was classified into three behavioral states: forward locomotion, backward locomotion, and stationary periods. For each frame, instantaneous velocity vectors were calculated from the change in centroid position between consecutive frames. The orientation of the movement was determined by the vector from the body centroid relative to the head position. When centroid direction was towards the head, movement was classified as forward; centroid direction away from the head was classified as backward movement, while low-magnitude velocities (less than 0.6 pixels per frame) were classified as stationary behavior. To ensure accurate movement classification, some corrections were applied. These arose during stationary and turning periods, where head detection was less accurate. Frames with movement classifications inconsistent with surrounding temporal context were identified and corrected. Continuous periods of each movement type were identified as behavioral bouts. For each bout, both temporal duration (in frames) and spatial extent (total distance traveled) were calculated. Movement metrics were calculated for each analyzed video, including, for example, total path length, maximum displacement from origin, proportion of time spent in each behavioral state, and average bout duration for forward and backward movement.

Microfluidic imaging of sensory-evoked calcium dynamics in RIA.

For recordings of RIA interneuron calcium dynamics, the strain MMH99 (pglr-3a::GCaMP3.3) was used. The worms were grown under standard conditions at 20°C, fed OP50 E. coli bacteria and assayed at the early adult stage [52].

Time-lapse fluorescence imaging of the RIA interneuron in living semi-restrained animals was performed in a microfluidic device as described in Hendricks et al. [53]. The aforementioned MMH99 strain selectively expressing GCaMP3.3 in RIA was used to measure calcium dynamics within the three distinct axonal compartments of RIA. Streams of liquid alternating between neutral NGM buffer and the attractive odorant IAA (isoamyl alcohol) were passed in front of the animal’s freely moving head to elicit sensory-evoked calcium events [54]. 512 x 512 TIFF stack recordings were acquired for individual animals over 1 minute at 10 frames per seconds and 100 ms exposures on an Olympus IX83 inverted microscope using a 40x U-PlanS Apo (N.A. 1.25) silicone immersion objective, Hammamatsu Orca Flash 4.0LT SCMOS camera, with the cellSens software (Evident). Automatic stimulus stream switches every 10 seconds were controlled by a solenoid pinch valve array (Automate Scientific).

Manual segmentation of axonal compartments and head bending extraction.

The brightness of the RIA compartments and the worm’s head position were manually extracted for comparison using the Fiji software as described in Ouellette et al. [32].

Automated segmentation of axonal compartments and head bending extraction.

Full-frame videos were processed to extract a region of interest centered on the RIA interneuron. Like for the other modalities, the Segment Anything Model 2 (SAM2) video predictor was initialized using the Hiera Large checkpoint. For each video, a single annotation point was placed on the RIA region in the final frame. This annotation was propagated backwards through all frames. The center of mass for each RIA mask was calculated across all frames to create a fixed crop window of 110 × 110 pixels around the RIA region throughout the video sequence. To simultaneously segment the three anatomical structures within the cropped RIA region (nerve ring dorsal (nrD), nerve ring ventral (nrV), and loop), a prompt management system was developed to enable consistent segmentation across videos. Expertly annotated prompt frames were incorporated into the video sequences to guide the segmentation model. SAM2 was reinitialized with the augmented video sequences, and structure-specific prompts. A comprehensive quality analysis was performed to identify segmentation artifacts, including empty masks, oversized regions, and overlapping structures. Additional prompt frames were added to the prompt pool if refinement was necessary. On rare occasions, sporadic frames with a missing mask were filled by interpolation. Occasional small noise artifacts were removed through connected component analysis. Brightness measurements were extracted from the segmentation masks by extracting the pixel intensities from the corresponding grayscale images.

Background correction was performed by sampling 100 background regions located at least 40 pixels away from any segmented object, as determined by distance transform analysis. The mean background intensity was subtracted from object measurements to obtain corrected brightness values. Since animals were positioned either on their left or right side in the chip, the spatial relationship between the loop structure and nrD was analyzed to adjust head bending angles as dorsal or ventral.

Full-body segmentation was performed on the original full-frame videos to enable postural analysis. A single annotation point was placed on the organism body and propagated through all frames using SAM2. The resulting masks captured the complete body, which served as input for subsequent head angle analysis. Morphological skeletonization was applied to whole-body masks to extract one-pixel-wide centerlines representing the organism’s medial axis. The angle between head and body vectors (anterior and posterior portions of the skeleton) was computed for each frame. Temporal smoothing was applied to remove measurement noise while preserving genuine behavioral transitions. A small temporal sliding window smoothing of 3 frames was applied to remove noise from small variations in body segmentation outline. Final head angle sign was adjusted based on the organism’s orientation determined in the previous step.

Human head angle measurement.

50 frames were randomly selected across the calcium imaging dataset described above. For each frame, a human expert measured the head angle by manually identifying the head tip and a straight reference line along the body midline using the angle measurement tool in Fiji [33]. The same 50 frames were fetched from both the Fiji ellipse-fitting and the TWARDIS pipeline outputs. Angles were orientation-corrected and translated to match a unified 0–180° convention, where 90° represents a straight head. Absolute differences from the human-measured angle were computed for each automated method.