Segmentation of lizard claws from computed tomography (CT) is challenging because claws are materially heterogeneous, containing both bone and keratinous tissue. Accurate tissue-level segmentation supports quantitative morphological and comparative anatomical analyses. This study develops and evaluates a deep learning framework aimed at automated 3D segmentation of lizard claw tissues from CT data, comparing architectures that either exclude or incorporate inter-slice spatial context.
The dataset comprised CT scans of 14 lizard claws. Reference segmentations (ground truth) were produced using a superpixel-based labeling workflow. The annotated dataset served both for training and for evaluation under a leave-one-out cross validation (LOOCV) regime, allowing each specimen to be held out in turn for validation.
The implemented segmentation pipeline used DeepLabV3 as the core segmentation architecture with three different backbone networks: ResNet-50, ResNet-101, and Inception-ResNet-v2. Two modes of input context were evaluated:
The study explicitly contrasted the performance of the 2D variants against the 2.5D variants to assess the influence of inter-slice information on volumetric segmentation accuracy.
Although model predictions were generated on a per-slice basis (2D or 2.5D inputs), outputs were reconstructed into three-dimensional volumes to permit volumetric comparison with the superpixel-derived ground truth. This reconstruction step enabled the calculation of 3D evaluation metrics and supports downstream morphological measurements.
Model performance was assessed using a leave-one-out cross validation (LOOCV) approach across the 14-claw dataset. Volumetric and boundary-based metrics reported included:
These metrics were used to quantify overlap, boundary agreement, detection sensitivity, and volumetric accuracy between predicted segmentations and the ground-truth labels.
Across all evaluation metrics the authors report that 2.5D CNN models consistently outperformed 2D counterparts, indicating a clear benefit from incorporating inter-slice contextual information for volumetric tissue segmentation. Among the tested backbones and input configurations, the 2.5D Inception-ResNet-v2 produced the best overall performance. The manuscript reports a validation accuracy of 97.5% for this model and states that its segmentations closely align with the ground-truth tissue structures.
Collectively, these results demonstrate that two-dimensional slice-wise segmentation benefits substantially when limited volumetric context is provided (2.5D), improving accuracy on heterogeneous anatomical structures such as lizard claws.
The study concludes that 2.5D deep learning approaches using DeepLabV3 with powerful backbones can deliver high-accuracy automated segmentation of heterogeneous lizard claw tissues from CT. The authors propose the pipeline as a robust framework for automated morphological analysis in comparative anatomy, where reproducible tissue segmentation can accelerate measurement and analysis of material heterogeneity in small anatomical structures.
As a preprint, the work has not undergone peer review. The manuscript describes experiments on a dataset of 14 claws and evaluates performance via LOOCV; further validation on larger and more diverse datasets would be required to confirm generalizability. Details beyond those reported — for example, precise training hyperparameters, computational requirements, and runtime performance — were not specified in the abstract and would need to be consulted in the full preprint or supplementary materials for reproduction.
This preprint was posted on bioRxiv on July 28, 2026. Authors are Haleema Sadia, Kate Manges Douglas, Asyiah Bray, Andrea Rummel, and Parvez Alam. The work lists funding from Edinburgh-Rice Strategic Collaboration Awards. The authors declared no competing interests. The preprint is available under a CC-BY 4.0 license. For correspondence the contact reported is parvez.alam@ed.ac.uk.