Nearly all specimens sampled to date in the bivalve-3D project (“biv3d”) are from museum collections. Sampling strategy will vary according to the arrangement of collections, but the following protocol has been applied in a variety of settings with no loss of specimens. Pulling and preparing specimens for scanning is best practiced with a joint physical and digital paper trail. From a given lot, the selected valve(s) are separated from the remainder of the lot and arranged for a photograph with the lot tag (Figure 1A). A high visibility tag (e.g., neon colored) is placed in the lot noting how many specimens have been pulled. Specimens are then prepped for transport and scanning depending on their size. If larger than 1 cm, the specimen is wrapped in low-density polystyrene foam (often sold as “dish wrapping foam”), which is secured with painter’s tape; both materials are transparent to X-rays so that specimens need not leave their packing for scanning, which greatly reduces risk of loss or breakage (Figure 1B). Crumpling the dish foam by hand makes it more pliable, helping to wrap more delicate specimens. Both the specimen’s registration number and its physical location are recorded with a pen on the painter’s tape. If the specimen is smaller than 1 cm, it is placed inside a gelcap and carefully secured in place with foam. The registration number and the physical location of the specimen (e.g., floor, row, cabinet, and drawer number) in the collection are written on a piece of paper and placed inside the gel cap.

At a minimum, we find that databasing the information in Table 1 is crucial to maintaining unique object identifiers and their associated metadata. It is important to record verbatim copies of the ID, locality and stratigraphy info as provided on the museum labels in order to maintain connections back to the museum database, even if those pieces of information are updated for analyses (see “lot photo” in Figure 1A). Each specimen picked for scanning receives a specimen ID (Table 1). In the case of specimens where both valves are to be scanned, a specimen ID refers to both valves, as they are part of the same specimen. This is necessary so that the two digital mesh objects representing those valves can continue to be associated in analyses by their specimen ID (each unique valve is then referenced by a mesh ID, Table 1).

For all scans, the size of the smallest feature of interest determines how specimens are grouped for bulk scanning. In bivalves, a key taxonomic character—the hinge teeth—are often an order of magnitude smaller than the shell, which sets an upper limit on how many specimens may fit into the field of view of the detector. For example, on the GE Phoenix v|tome|x M 240/180 kV Dual Tube μCT with a DXR250 (16″, 4 M pixel) X-ray detector at the U.S. National Museum of Natural History (NMNH, also USNM), specimens with approximately 1 mm (1,000 μm) features of interest are sufficiently resolved using <20 μm resolution, which would characterize such features with 50 pixels in the image plane. This resolution translates to a 40 mm by 40 mm field of view on a detector 2000 × 2000 pixel detector, which can typically accommodate 4–5 approximately 10–20 mm specimens in a cylindrical arrangement. Larger specimens have larger sizes of their smallest features of interest and can be packed to similar numbers for scanning at lower resolutions; however, most bivalves have scanned best at resolutions finer than 50 μm. Fossil specimens embedded in sedimentary matrix are often scanned alone because multiple fossil specimens, particularly those in dense siliciclastics, can require higher voltage to penetrate the full object; this may exceed the power limits of the micro-CT, and it will likely reduce the contrast between the edges of the specimens and matrix, making segmentation more difficult or virtually impossible (advances in laboratory phase contrast CT could improve segmentation of materials with subtle differences in signal, see Birnbacher et al., 2021). Additionally, scanning fossils embedded in matrix at the highest possible resolution reduces artifacts such as partial volume effects, which arise as the averaging of gray values from materials with different densities in a single voxel (described further in Abel et al., 2012; Racicot, 2016).

Once specimens have been grouped for scanning, they can be mounted into cylindrical containers. Straight-sided, thin-walled plastic soda bottles with a flat, level foam insert supporting specimens from the bottom work best for holding specimens (Figure 2A). For macro specimens (>1 cm), arrange in an imbricated fashion with the shell commissure perpendicular to the base of the container (Figure 2A, i.e., mounting with the long-axis perpendicular to the X-ray source as in Sutton et al., 2014, p. 53); this arrangement reduces Feldkamp artifacts and provides the sharpest boundaries between surfaces. Specimens can be mounted in vertical layers up to a height equaling the diameter of the container (i.e., a square field of view), and packed with additional foam to prevent movement during scanning. For micro specimens (<1 cm), use paper or plastic drinking straws to hold gel caps; straws can then be inserted into 50 ml or 15 ml conical centrifuge tubes (Figure 2B), making sure to keep specimens level with each other.

To orient the scan, place a marker into the scanning container with lower X-ray density than the specimens, such as an eraser; using markers that are more X-ray dense can shade specimens, which makes them more difficult to segment in post-processing. Draw a plan view “scan map” of the specimens relative to the marker, identifying positions using the unique identification code labeled on the specimen wrap or gel cap (Figure 2). Each scanning “cartridge” can be prepared before scanning, which is a crucial step for maximizing machine time.

Once specimens are mounted in the micro-CT, several parameters should be tuned to optimize the X-ray imaging. Each scan can, and often does, have different parameter settings, which vary according to the material density of the specimens and the number of specimens in the scan scene—or, for fossils, by the volume and mineralogy of the surrounding matrix. Because specific scanning parameters will vary by micro-CT system, we emphasize that the tuning process outlined below is intended to be generalizable; nevertheless, the parameters reported here are likely good starting points (for a comprehensive overview of optimizing scanning parameters, see Sutton et al., 2014, pp. 56–60). For all steps that follow, monitor the histogram of gray values in the imaging software; this is the crucial tool for maximizing contrast between materials in the scan.

Because we aim for high throughput, we first set parameters that affect the length of the scan to their lowest values: i.e., set exposure timing to lowest value, frame averaging to 1, and frame skips to 0. Next, we rotate the mounted specimens until viewed through the thickest portion of the scan scene; this will correspond to the lowest intensity gray values in the live image. We increase or decrease the voltage until the minimum gray value intensity recommended for the machine is reached (this ranges between 100 and 200 for the micro-CT at NMNH, trending towards the higher end as the filament ages). We aim for the lowest energy X-rays necessary to penetrate the specimen, which maximizes the contrast between materials (Sutton et al., 2014, pp. 56–57). Next, we adjust the current to maximize the total range, contrast ratio, and signal to noise of the histogram (Table 2). Increasing the current also increases the power, and tends to increase the minimum gray value, which can reduce both the signal to noise and contrast ratios (compare values in Table 2); if so, step back to previous settings. If the minimum gray cannot be reached within the power limits of the X-ray tube or without oversaturating the detector, add a physical filter (see next section) or increase the exposure timing, which may require adding frame skips so that the detector panel can discharge between images (this timing can be highly system-dependent). Similarly, it is important to check that the X-ray spot size is not 1.3 times larger than the expected voxel size (i.e., resolution) of the scan. Increasing exposure time and frame averaging will also increase the signal to noise ratio, helping to define sharper boundaries between shell and matrix if needed (often a good setting to adjust when the shell is a calcareous matrix).

Filters can reduce beam hardening, an artifact where soft X-rays are absorbed at the sample surface leaving only the higher energy X-rays to penetrate the sample; this creates streak or cupping artifacts, where gray values grade from high to low intensity towards the interior of the specimen (see Sutton et al., 2014; Wellenberg et al., 2018). Fossil shells in matrix frequently show cupping artifacts, which can complicate simple segmentation on gray value intensity. Metal filters, typically aluminum, copper, or tin sheets ranging from 0.1 to 1 mm in thickness, can minimize this gradient in gray value intensities, but at the cost of diminished compositional contrast. Because many of our fossil shells are in calcareous matrix, meaning there is minimal compositional contrast between specimen and rock, we generally do not use filters in order to maximize what little contrast may exist. For Recent specimens free of matrix, a 0.1 mm copper filter does reduce beam hardening, simplifying segmentation of shell material from air and foam, but segmentation of scans lacking a filter is not any more difficult.

All of our scans use 360° rotation, and while the rule of thumb for setting the number of projections images is the width of the scan in pixels multiplied by π (Keklikoglou et al., 2019, p. 17), we find that shells with simple, large morphologies (i.e., smooth shells with large hinge teeth) can be scanned with projections equaling the greatest width of the scan scene (for a detector 2000 pixels wide, this would be 2000 projections). This can reduce scan times to as low as 7–10 min, depending also on the exposure time, frame averaging and skips. Thus, it is important to group specimens for bulk scanning according to the desired resolution for their smallest features of interest. For shells with relatively finer details, or for those embedded in matrix, setting a higher number of projections can better resolve those features (Sutton et al., 2014, p. 58; Keklikoglou et al., 2019, p. 7), although we find N*1.5 to usually be sufficient.

We have reconstructed all scans using the proprietary software of the micro-CT systems (i.e., GE phoenix datos|x), but alternative solutions are available (Sutton et al., 2014, pp. 149–150). We have always scanned with settings to compensate for any drift in the center of rotation, changes in focal spot size, and any small movement of the specimens (in the GE system, this would be using the Autoscan Optimizer followed by the Automatic Geometry Calibration before reconstruction). Scanning using detector shifts (i.e., small adjustments in the detector panel to reduce hot pixels), can reduce beam hardening (for additional mitigation techniques, see Wellenberg et al., 2018). Streaks, ring artifacts, and beam hardening are often in our scans and could be further mitigated during the reconstruction step—but we find these artifacts to rarely impact the final surface mesh for Recent and fossil material that is free of matrix. For fossils in matrix, application of deep learning segmentation below can help to overcome these artifacts.

Once reconstructed, Recent and fossil material that is free of matrix can often be isolated from air and packing materials by simply thresholding out the non-target values in the image stack (see a recent list of commercial software and freeware for this operation in Keklikoglou et al., 2019, their Table 15, but also see 3D Slicer Kikinis et al., 2014; we primarily use ORS Dragonfly; Object Research Systems (ORS) Inc., 2022). If the organic content of shells varies within a bulk scan (i.e., nacre vs. calcite), the scan scene may require separate threshold values for generating shell-specific regions of interest (ROIs). These shell ROIs can be further refined using manual segmentation tools, or by removing any pixels isolated from the targeted shells with simple region of interest tools, such as island processing algorithms that remove pixel sets less than a certain value (e.g., “Processing Islands” in ORS Dragonfly). From these refined ROIs, contour meshes can be generated and exported for further cleaning and morphological analysis. For many fossil shells embedded in matrix, particularly those in calcareous matrix, this simple thresholding approach will not work given the limited compositional contrast between the specimen and sediment. Here, deep learning can facilitate segmentation.

Deep learning has become a powerful tool for image segmentation, especially for digital excavation of fossils embedded in matrix (Liu and Song, 2020; Yu et al., 2022). However, we note that to date, our applications of deep learning have yet to perfectly segment a bivalve fossil from matrix (as Yu et al., 2022 also noted for segmenting dinosaur bones from matrix). Thus, we use deep learning as a technique to speed up segmentation, which almost always requires manual tuning to complete; this means finding a compromise between the time spent fitting the model(s) and then manually finalizing the segmentation. We use the deep learning interface developed by ORS Dragonfly (the “Segmentation Wizard,” Badran et al., 2020), which creates an interactive session for fitting and refining multiple deep learning models for segmentation. Regardless of the software used, we suggest that this approach is a good, generalizable workflow. Continued research into model architectures will undoubtedly bring faster and more accurate initial segmentations of these low compositional contrast materials, but our emphasis here is on producing sizable datasets of workable 3D models for quantitative evolutionary analysis. Thus, the trade-off in time spent tuning hyperparameters and fitting models in search of a perfect segmentation should be weighed against using an adequate model to provide a strong starting point for manual segmentation.

Within the ORS Dragonfly Segmentation Wizard, we begin segmentation by selecting a slice from the image stack that best represents the full diversity of features; often this is an image slice containing hinge teeth or fine ornamentation (e.g., Figure 3A). We then define the segmentation frame (i.e., a rectangular mask) that brackets the shell and manually set ROIs corresponding to shell material and background (i.e., matrix and/or air, Figure 3A, bottom row). We then train a series of deep learning models using this single frame and visually inspect the predicted segmentation (Figure 3B). Usually, the U-Net architecture (Ronneberger et al., 2015) is the fastest and most accurate approach, but Sensor3D, and variations on the U-Net such as Attention U-Net and U-Net++ can sometimes better separate the boundaries between fossil shell and matrix (hyperparameters for the U-Net fit here are in Figure 3C). The resulting model accuracies provide a general estimate of segmentation performance, but high accuracies can sometimes characterize models with poor definition of the boundary between fossil shell and matrix (e.g., frame 1 in Figure 3B for Round 1 of training, which had an accuracy score of 0.97 Figure 3D). In this example, the model trained for one round cleanly separates shell from air, but confuses parts of the matrix and shell (compare Figures 3A,B for frame 1, particularly around the hinge area). Therefore, the additional training data was labeled, and the model weights updated through continued training.

Subsequent rounds of training improved the segmentation of shell from matrix (compare better capture of the hinge area in rounds 2–4, Figure 3B). The training data in frames 3 and 4 were selected to improve segmentation of shell from matrix where streak artifacts impacted the shell boundary at the edge of the specimen. Frame 5 was selected to improve segmentation of shell from matrix for a region of the hinge area that was not well predicted through the third training round. We could continue to iterate the process at this point, expanding the training set to include image slices with poor predicted segmentation. However, as mentioned above, we weigh the continued time in fine-tuning these segmentation models against manually tuning the segmentation. In general, we find that one or two rounds of creating training data and fitting models is sufficient to produce a strong starting segmentation. From there, we manually segment regions where the model failed to define the boundary between shell and matrix. Without any manual segmentation, the resulting 3D model has noise, but captures nearly all of the relevant taxonomic details (Figure 4). Timing required for segmentation via deep learning compared to fully manual operation depends, in part, on compute power. The process described for the fossil in Figure 3 required approximately 3 h, from labeling training data to fitting the model and predicting the segmentation, and half of that time was fitting the initial model (compute power: Intel Xeon Silver 4214R CPU @2.40 GHz with 512 GB RAM and 2 × 16 GB NVIDIA Quadro RTX5000 GPU). Fully manual segmentation of this fossil is estimated to take 8 or more hours of constantly engaged work.