Brain virtual histology and volume measurement of a lizard species (Podarcis bocagei) using X-ray micro-tomography and deep-learning segmentation

Introduction

There has been an increasing interest in studying the individual variation in animal behaviour and cognition and understanding the underlying neural correlates (Gonda, Herczeg & Merilä, 2013; Thornton & Lukas, 2012; Triki & Bshary, 2022; Triki et al., 2019). Understanding intraspecific variation in the brain and linking it to the individual-level ecology, like social group size, dominance hierarchy, foraging behaviour, mating strategies, and predator avoidance (Guadagno & Triki, 2023; Kolm et al., 2009; Triki et al., 2020, 2019; Vega-Trejo et al., 2022), among other selective pressures, can help us to understand better brain development and, ultimately, brain evolution.

Ectotherms, including fish, amphibians, and reptiles, represent a promising vertebrate group for exploring the relationship between ecology, cognition, behaviour, and the brain. Their ability to maintain neurogenesis in adulthood and generate neural tissue when necessary for adaptation enables the detection of variation within species and even within populations (Gonda, Herczeg & Merilä, 2013; Zupanc, 2006). However, several challenges emerge when studying vertebrate brain morphology and its relationship to ecology (social and environmental conditions). For example, to separate species-specific effects and account for phylogenetic signals, comparative brain studies require extensive sample sizes and a diverse array of species. This is difficult due to the labour-intensive methods needed to collect brain morphology data, which often results in studies with small sample sizes. Additionally, in comparative research, it is essential to standardise methodologies and reduce observer bias (Tuyttens et al., 2014).

A plausible solution involves measuring total brain size, which seems relatively straightforward, as it necessitates fewer complex techniques to ascertain volume or weight. Additionally, it is recognised as one of the most extensively studied traits in evolutionary biology, both in existing and extinct species (Isler et al., 2008). Acquiring the size of different brain regions complicates the matter. In neuroecology, cost-effective techniques such as ellipsoid measurements are frequently used to estimate the sizes of brain regions in several teleost fish species from lateral, ventral, and dorsal brain pictures, e.g., cichlids, guppies, and gobies (Pollen et al., 2007; Triki et al., 2023; White & Brown, 2015). The technique uses pictures of brains from lateral, ventral, and dorsal sides and estimates the volume of each brain lobe from its length, width, and height, fitting these dimensions into the ellipsoid formula: volume = (length × width × height) × π/6. While the method offers useful insights, it has low accuracy by simply assuming that all the lobes are ellipsoidal in shape (White & Brown, 2015). Consequently, the ellipsoid method is not generalizable across species, particularly when brain shapes differ markedly from an elliptical form. Ectothermic species typically have long, narrow, and tubular brains, with a more linear organization in which different regions are arranged sequentially. Their brain lobes are often more ellipsoidal in shape, which may align more closely with the assumptions of the ellipsoid method. In contrast, endothermic brains tend to be more compact and rounded, making the method less suitable for these species (Striedter, 2005).

Another method widely used for estimating brain region sizes is the histological technique. It involves sectioning the brain into slices, then staining and estimating the surface area of each region on a brain slice, summing these areas, and multiplying them by the thickness of each slice (White & Brown, 2015). The histological technique estimates various lobe volumes more precisely than the ellipsoid method. Additionally, it considers the irregular shapes of lobes and the beginning and end of ventricles, aspects that the ellipsoid estimation cannot account for. Nonetheless, the histological method has significant limitations, mainly because obtaining brain region measurements is extremely time-consuming and tissue-destructive (Cnudde et al., 2008; Johnson et al., 2006; Kuan et al., 2020), rendering it less appealing than the ellipsoid technique.

In the past decade, modern imaging methods, previously reserved for medical usage and clinical research, have made a significant impact on neuroecology, such as magnetic resonance imaging (MRI) and X-ray tomography. These methods provide 3D images of the internal structures in a non-destructive way, allowing accurate volumetric measurement, as well as repeatability and sample reusability, which is especially important for wild specimens and animal welfare purposes.

So far, two studies have provided 3D models of lizard brains, the tawny dragon (Ctenophorus decresii) and the Bearded Dragon (Pogona vitticeps), using MRI (Foss et al., 2022; Hoops et al., 2018, 2021). In this study, we provide a detailed step-by-step protocol (Zhou, Dragunova & Triki, 2024) for extracting 3D brain morphometric data using X-ray micro-tomography (microCT) from a common lizard, Podarcis bocagei. To increase efficiency and reduce time-consuming sample preparation, we prepared whole heads instead of dissecting the brains. This also avoids potential damage during dissection, embedding and sectioning as in traditional histological methods.

Using microCT 3D data and deep-learning neural network (Lösel et al., 2020; Ronneberger, Fischer & Brox, 2015), we anatomically segmented six distinct major brain regions: olfactory bulbs, telencephalon, diencephalon, midbrain, cerebellum, and brain stem (Bruce, 2009) of 29 specimens. Deep learning segmentation for 3D images is fast growing in medical imaging (Liu et al., 2021), where the training data size is generally in the range of hundreds (Oktay et al., 2018). It is uncommon to have such large datasets for studies on wild animals. However, our study demonstrated that deep learning segmentation could yield satisfactory results with as few as five training sets (brains), which has the potential to significantly reduce manual segmentation efforts for both large and small-scale studies. Portions of this text were previously published as part of a preprint (Zhou, Dragunova & Triki, 2024).

Materials and Methods

X-ray microCT scan

The samples were scanned using Zeiss Xradia Versa 520 at Stockholm University Brain Imaging Centre, with the samples being stabilised inside 5 mL plastic tubes and remained in the staining solution, as shown in Fig. 1. The X-ray source was set to have a voltage of 100 kV and a power of 9 W. The 0.4x objective and a charge-coupled device (CCD) camera were coupled with a scintillator. The effective voxel size was 17.4 μm, with the optical and geometrical magnification being compensated. The scan consisted of 801 projections over 212 degrees with 1 s exposure time for each projection. In total, one scan took 36 min, including reference images and readout time of the CCD camera. Autoloader was utilised to change samples automatically to minimise the manual work. The tomography reconstruction was done automatically with Zeiss Scout-and-Scan software right after the scans, and the output was 16-bit grey value tiff image stacks.

Semi-automatic segmentation of training datasets

It was not possible to manually place all samples in exactly the same position with an accuracy of tens of microns. For this reason, the first step in preparing the data were to align and crop all the images, as demonstrated in Fig. 2A. This was performed using the software Dragonfly. This step may be overlooked but it was crucial for a consistent brain volume measurement, especially the brainstem, and later for efficient neural network training.

The training dataset was segmented utilising semi-automated methods and was comprised of three distinct steps. The first step was manual segmentation of approximately 20–30 slices in total of the three planes in Dragonfly (Figs. 2B, 2C). The second step was random walk interpolation in Biomedisa (Lösel & Heuveline, 2016; Lösel et al., 2020). The third and final step was checking the segmentation and correcting manually any inconsistencies if needed. As the boundaries of certain brain regions can be indistinct in some areas, we have manually delineated these brain regions in the initial step of the segmentation, following brain anatomy descriptions and brain atlases from lizard species (Bruce, 2009; Hoops et al., 2018; Menezes Freitas, Paranaiba & Lima, 2023). A sketch of the procedure from semi-automatic labelling to deep-learning segmentation is shown in Fig. 2D.

Results

Virtual histology and 3D model from X-ray micro-tomography (microCT)

Figures 3 and 4 display examples of virtual histology sections from a single brain in coronal, sagittal, and horizontal orientations. Several brain areas are labelled with references from Hoops et al. (2018), Menezes Freitas, Paranaiba & Lima (2023), Naumann et al. (2015) (abbreviations in Table 4). The brain virtual sections in Figs. 3 and 4 were generated with a mask from segmentation without manual brain dissection. The results also provide virtual histology of the whole head, offering detailed information about tissues beyond the brain. For example, in Fig. 5A, the olfactory nerves can be identified, and the diameter is approximately 26 μm as shown in the inset.

Table 4:

Abbreviations of brain regions.

Abbreviation	Brain region
4V	Fourth ventricle
AOB	Accessory olfactory bulb
aon	Accessory olfactory nerve
Ce	Cerebellum
DC	Dorsal cortex
DMC	Dorsomedial cortex
DVR	Dorsal ventricle ridge
GL	Granular layer of the cerebellum
LC	Lateral cortex
lfb	Lateral forebrain bundle
LV	Lateral ventricle
MC	Medial cortex
ML	Molecular layer of the cerebellum
MOB	Main olfactory bulb
on	Olfactory nerve
OT	Optic tectum
ot	Optic tract
TV	Tectal ventricle

DOI: 10.7717/peerj.19672/table-4

Brain region segmentation

Using deep learning methods, we segmented six main brain regions from 29 specimens, namely, the olfactory bulb, telencephalon, diencephalon, midbrain, cerebellum, and brainstem (see Fig. 6) (also see Video S1). The average volume values of all 29 segmented brains are reported in Fig. 7 using MATLAB, with the standard deviation as the error bars (see Table S1).

The results of using both deep learning methods, Biomedisa and AIMOS, are compared and presented in Fig. 8 (Table S2). We found that, because AIMOS is a 2D approach, the orientation of the slices influenced the results (Fig. S2). We identified that the sagittal plane works best for AIMOS and therefore chose it for the comparison with Biomedisa. In contrast, Biomedisa operates on a 3D representation and is not restricted to a specific data plane orientation. In Fig. 8, five separate networks were trained for each algorithm using three, five, seven, nine, and 11 images, where each trained neural network was then used to predict four test image stacks. The predicted segmentations were evaluated using the average dice score (DSC) (Bertels et al., 2019) and the average relative error of brain region volumes (RE). It can be seen that as few as three samples provided sufficient accuracy with DSC > 0.9 using Biomedisa. As AIMOS required two samples from the labelled data for validation, the input data size cannot be smaller than three. For both AIMOS and Biomedisa, RE was about 4–5% when the training sets were 5 or more, as shown in Fig. 8 (Table S2).

Overall, comparing the two algorithms revealed a visible difference for small data sizes. At an input size of 3, Biomedisa far outperforms AIMOS. Nevertheless, when the input data size exceeds 5, this difference in performance disappears.

Discussion

Twenty-nine specimens of common lizard Podarcis bocagei were scanned using X-ray microCT to generate 3D images with internal structures in high resolution. Six brain regions, olfactory bulbs, telencephalon, diencephalon, midbrain, cerebellum, and brain stem, were segmented, and their volumes were measured. Two deep learning segmentation algorithms were applied and compared quantitatively. The results have shown that as few as five training datasets (five samples) were sufficient for both algorithms. Biomedisa operates in 3D and is not restricted to a distinct plane orientation. AIMOS uses 2D U-Net, the orientation of the slices influenced the results. For AIMOS, the sagittal plane works best in this study as sagittal slices contain most of the brain regions simultaneously.

The methods described in this study offer an efficient protocol for achieving 3D imaging acquisition and volume measurement with minimal manual work and high accuracy in several aspects. First, it eliminates manual work in brain dissection and sectioning like in classical histological techniques. Second, it offers more accurate volumetric measurements compared to the ellipsoid method (White & Brown, 2015). Third, it significantly reduces manual labelling time by leveraging smart interpolation and deep learning prediction. Instead of manually segmenting approximately 200 slices from 29 samples, we trained the model using just 30 slices from five samples—achieving a 39-fold speedup. This translates to 2 weeks of work with our protocol compared to 78 weeks (18 months) of manual labelling. Finally, it increases data accessibility by facilitating the reuse and reproducibility of the findings.

In neuroecology, there are several limitations to conducting research that links brains to behaviour and cognition, especially in wild animals. Studying wild animals and obtaining sufficiently large sample sizes can prove challenging (Bshary & Triki, 2022). To address this, researchers often tend to maximise the amount of data collected from wild specimens, both to compensate for small sample sizes and to prioritise animal welfare (Bee et al., 2020). Therefore, creating a database with 3D brain scans beyond the scope of individual studies would be highly beneficial, as it would enable other researchers to access these data and explore new questions related to different traits. For instance, in addition to measuring the size of brain regions, high-resolution 3D scans can provide more comprehensive data on other traits, such as those associated with vision (e.g., Fig. 5B). Compared to traditional histological techniques, these scans can offer valuable insights without the risk of deformation during dissection. Accurate 3D modelling is critical for simulating the optical properties and visual systems in various species, including humans, by extracting the 3D morphology of eye structures (MercuȚ et al., 2020; Taylor et al., 2020). This enables precise estimation of key parameters such as resolution, focal length, and field of view. Moreover, other anatomical information, such as tooth development, skull structure, and jaw anatomy, can also be gathered from these scans (see examples in Figs. 5C, 5D) (Ballell et al., 2024; Beazley et al., 1998; Urošević, Ljubisavljević & Ivanović, 2014; Zahradnicek, Horacek & Tucker, 2012).

While X-ray microCT may not match the 2D resolution of histological images, it offers isotropic resolution in three dimensions, providing more detailed information in the sectioning dimension and enabling virtual histology in any direction from the same scan. This reduces the need for manual sample sectioning and minimises the number of animals required, making X-ray microCT an ideal tool for volumetric studies. MicroCT is not as cost-efficient as ellipsoidal volume estimation, but it is less costly than MRI. Additionally, compared to MRI, X-ray microCT provides higher resolution and better contrast for bones, making the data potentially useful for other studies. The staining protocol in this study can be adjusted according to the requirements of the studies, for example, iodine staining can be used for faster staining, but it is not suitable for long-term storage due to the destaining (Callahan et al., 2021). In contrast, for museum specimens or other samples that need to be destained afterwards, extra steps, including NaOH washing, are required if PTA is used (Hanly et al., 2023).

Importantly, the segmentation protocol outlined here is not limited to X-ray microCT data; it can be applied to any 3D scans with different training datasets tailored to various imaging modalities. Additionally, we have made our 3D microCT data, along with the corresponding labels and trained neural network, freely available (Zhou, Dragunova & Triki, 2025).

With the growing number of studies in computer science highlighting the potential of deep learning for various image processing tasks, this study aims to provide a practical and easy-to-follow guideline. We hope this protocol will be valuable to researchers across diverse fields beyond ecology and evolution, enabling faster, more reliable, and reproducible brain segmentation. Computer scientists could also benefit from accessing our data to develop more efficient algorithms for various tasks, as they are currently working on constructing algorithms to “segment anything” for general 2D images (Kirillov et al., 2023), as well as for 3D medical imaging (Ma et al., 2024). Without a sufficient available data bank, the development of algorithms for animals other than standard model organisms would not be possible. Future work can focus more on improving the training efficiency, combined with data augmentation and simulation, to push the limits of training dataset size further and bring manual annotation efforts to a minimum.

Supplemental Information