pith. sign in

arxiv: 2605.15093 · v2 · pith:2MUFJOH3new · submitted 2026-05-14 · 💻 cs.CV

CoralLite: {μ}CT Reconstruction of Coral Colonies from Individual Corallites

Jess Jones , Leonardo Bertini , Kenneth Johnson , Erica Hendy , Tilo Burghardt This is my paper

Pith reviewed 2026-05-20 20:58 UTC · model grok-4.3

classification 💻 cs.CV
keywords coral skeleton reconstructionmicro-CT segmentation3D corallite modelingdeep learning for biologyPorites colony analysisV-Trans-UNettopology-aware training
0
0 comments X

The pith

A hybrid neural network reconstructs individual 3D corallites from micro-CT scans of entire coral colonies.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents CoralLite, a dataset of annotated micro-CT scans of Porites coral skeletons together with a baseline deep learning method for segmenting and linking individual corallites across slices into full 3D models. This capability would let biologists trace the skeletal record left by each polyp and thereby measure division rates and growth timing inside colonies that can live for centuries. The method uses a hybrid V-Trans-UNet pre-trained on weakly labeled data and then topology-aware fine-tuned on more than eight thousand manual corallite annotations from selected slices. On held-out slices from the training colony the model reaches 0.94 topological accuracy and a mean Dice score of 0.77; performance drops to a mean Dice of 0.63 on a biologically unrelated specimen, yet the authors conclude that visual machine learning can already support complete 3D corallite reconstruction from scans alone.

Core claim

CoralLite supplies the first fully quantified volumetric segmentations of entire calcareous coral skeletons with cross-slice linking that produces 3D visualizations of each corallite up to colony scale. A hybrid V-Trans-UNet architecture is pre-trained on weakly annotated data and topology-aware fine-tuned on 8k+ manual corallite region annotations from limited slice sections. On unseen slices of the same colony the model attains 0.94 topological accuracy at mean Dice 0.77; on a different, biologically unrelated specimen it attains mean Dice 0.63. These results demonstrate for the first time that visual machine learning can support full 3D individual corallite modelling from μCT scans of珊瑚骨骼

What carries the argument

Hybrid V-Trans-UNet architecture that segments tiled μCT virtual slabs of coral colonies after weak pre-training and topology-aware fine-tuning on manual corallite annotations.

If this is right

  • Scientists can now extract the division history of individual polyps from the skeletal record archived in massive colonies.
  • Cross-slice linking produces consistent 3D corallite models that can be visualized and measured at colony scale.
  • The published dataset of 697 slices, 37 annotations, network weights and code provides a reproducible baseline for further coral reconstruction work.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the segmentation generalizes, repeated scans of the same colony over time could reveal how environmental stress alters polyp division patterns.
  • The same slice-to-volume linking approach might be tested on other calcifying organisms whose growth is archived in layered skeletons.

Load-bearing premise

Fine-tuning on annotations from limited slices of a few colonies will produce reliable 3D reconstructions when applied to complete colony volumes and to biologically unrelated specimens.

What would settle it

Apply the trained model to complete μCT volumes of several unrelated coral species, count the resulting 3D corallites, and compare the count and topology against exhaustive manual expert annotation of the same volumes.

Figures

Figures reproduced from arXiv: 2605.15093 by Erica Hendy, Jess Jones, Kenneth Johnson, Leonardo Bertini, Tilo Burghardt.

Figure 1
Figure 1. Figure 1: Overview of CoralLite. The proposed pipeline comprises three main stages: (1) Novel Dataset: Sequential 2D cross-section slices from 3D µCT scans of a Porites sp. coral colony are annotated and processed into 5x224x224 tiled snippets. (2) Volumetric Segmentation: A 3D-context-aware Trans-UNet pipeline with a hybrid ResNet50-ViT backbone (details in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset Sample Slices next to Segmentation Results. Shown are 5 of the 697 full-colony slices from the dataset next to segmentation outputs using inference performed on the Naturalis 6781 colony. This specific coral colony was originated from a single coral polyp forming a single corallite. The coral grew over a five year period into a lumpy hemispheric colony of thousands of corallites. Depicted are sampl… view at source ↗
Figure 3
Figure 3. Figure 3: Volumetric Input Tensor Generation. Tiled volumetric stacking provides local context from adjacent slices for segmentation. Left to right: Illustrative example of raw image tiling with window 224×224 and step size 224; An example composite tile with annotation overlay in red, from the Naturalis 6785 dataset; An example snippet of dimensions 5 × 224 × 224. Note that slice n is situated in the centre of the … view at source ↗
Figure 4
Figure 4. Figure 4: Full Architecture Diagram. Our pipeline uses a Video-TransUNet [22] de￾sign but applied here to spatial volume analysis. It utilises CNN feature encoding with short-range volumetric learning and complex spatial context modelling via a ViT before DCNN decoding: (a): 5 × 224 × 224 input tensors are provided to (b): a pre-trained ResNet-50 CNN encoder with 3x skip connections. (c): Feature encodings are blend… view at source ↗
Figure 5
Figure 5. Figure 5: Topological Error Maps. Two example depictions of connectivity-driven error maps used for training loss calculation highlighting incorrectly connected or gen￾erated corallite regions. The topological error map E is presented on the left of each image pair, while composite ground-truth (blue) and prediction (red) maps are given on the right of each pair. The grayscale error map demonstrates how the topologi… view at source ↗
Figure 6
Figure 6. Figure 6: Topological Contribution Tuning, Training and Validation Plots. Performance development during main fine-tuning of our model. (left) A compari￾son of different topology coefficients T confirms stable performance around 0.1-1.0. (right) With peak validation accuracy of 0.96 and saturated loss reduction at 320 epochs at a stable plateau, the depicted model (FT VLrg 0.1xT + GN in Tab. 2) has learned to segmen… view at source ↗
Figure 7
Figure 7. Figure 7: Connectivity Impact of Topological Loss. Four composite example pairs of predictions (red) and ground-truth data (blue) superimposed on coral tiles, that is (left) without and (right) with topological loss utilised (k = 50, T = 1). Note the improvement in structural corallite separation, particularly in areas highlighted by the red bounding box. Bottom right pair: Note that scanning noise and xray echoes, … view at source ↗
Figure 8
Figure 8. Figure 8: Corallite Reconstruction and 3D Visualisation. (left) Per-slice corallite representation based on centroid, major and minor axis lengths, and in-plane orienta￾tion shown as vector pair. (middle) Corallite #87 as viewed along the X-axis where centroid tracing has translated stacks of 2D segments into a coherent 3D structure for visualisation that captures the natural geometry, curvature, split and terminati… view at source ↗
Figure 9
Figure 9. Figure 9: Further Segmentation Results. Examples from Naturalis 6781 colony; (left) Annotation example, slice 1279 (Growth). (right) Inference outputs, clockwise from top-right: slice 2043 (Ortho), 2718 (Ortho), 1379 (Growth), and 0119 (Growth). 7 Conclusion We presented CoralLite, a new µCT coral dataset, with individual corallite an￾notations and a V-Trans-UNet inspired deep learning pipeline for corallite seg￾men… view at source ↗
Figure 10
Figure 10. Figure 10: Example of Colony-scale Reconstruction (Close-up Visualisation). [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
read the original abstract

The life history of an individual coral is archived within the accreting skeleton of the colony. While reef-forming coral colonies (e.g. massive $\textit{Porites}$ sp.) may live for hundreds of years and deposit calcareous structures many metres in height and width, their living tissue is a thin outer surface layer comprised of asexually-dividing polyps that only survive a few years. To understand the rate and timing of polyp division and the consequences for colony skeletal growth, scientists need to track the skeletal corallite deposited around each polyp. Here we propose CoralLite, an annotated $\mu$CT scan dataset of entire calcareous skeletons and an associated, first corallite deep learning reconstruction baseline. CoralLite combines fully quantified volumetric segmentations with cross-slice linking for visualisations of 3D models for each corallite up to colony scale. For segmentation, we propose and evaluate in detail a hybrid V-Trans-UNet architecture applicable to segmenting tiled $\mu$CT virtual slabs of $\textit{Porites}$ sp. colonies. The model is pre-trained on weakly annotated data and topology-aware fine-tuned using fully annotated slice sections with 8k+ manual corallite region annotations. On unseen slices of the same colony, the resulting model reaches 0.94 topological accuracy at mean Dice scores of 0.77 on the same colony and projection axis, and 0.63 mean Dice scores on a different, biologically unrelated specimen. Whilst our experiments are limited in scale and context, our results show for the first time that visual machine learning can effectively support full 3D individual corallite modelling from $\mu$CT scans of coral skeletons alone. For reproducibility and as a baseline for future research we publish our full dataset of 697 $\mu$CT slices, 37 partial or full slice annotations, and all network weights and source code with this paper.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the CoralLite dataset consisting of 697 μCT slices from Porites sp. coral colonies along with 37 partial or full slice annotations totaling over 8k manual corallite regions. It proposes a hybrid V-Trans-UNet architecture that is pre-trained on weakly annotated data and topology-aware fine-tuned on the annotated slices. The model achieves mean Dice scores of 0.77 on unseen slices from the same colony/projection axis and 0.63 on a biologically unrelated specimen, together with 0.94 topological accuracy on held-out slices. The authors state that the approach, combined with cross-slice linking, enables full 3D individual corallite modelling at colony scale and claim this is the first demonstration that visual machine learning can support such reconstructions from μCT scans of coral skeletons alone. The dataset, annotations, network weights, and source code are released.

Significance. If the 3D reconstruction pipeline proves reliable on full colony volumes, the work would offer a practical baseline for coral biologists to quantify polyp division rates and skeletal growth patterns archived in the colony structure. The public release of the annotated μCT dataset and reproducible baseline constitutes a clear strength, providing a starting point for future computer-vision research in this specialized domain. At present, however, the significance is tempered by the modest scale (two specimens) and the absence of direct 3D quantitative validation.

major comments (2)
  1. [Abstract / Evaluation] Abstract and evaluation sections: the central claim that the method supports 'full 3D individual corallite modelling' rests on 2D per-slice segmentation metrics (Dice 0.77/0.63, topological accuracy 0.94 on held-out slices). No 3D metrics—volumetric IoU, corallite instance tracking precision across slices, or consistency over full colony height—are reported, leaving the cross-slice linking step and its reliability on complete volumes unquantified.
  2. [Abstract] Abstract: generalization is demonstrated on only two specimens, with a substantial drop to Dice 0.63 on the biologically unrelated specimen. The dataset comprises 697 slices and 37 annotated sections; the authors themselves flag the experiments as limited in scale. Additional colonies and quantitative assessment of 3D reconstruction on full volumes would be required to support the broader claim.
minor comments (2)
  1. [Methods / Experiments] The manuscript does not report error bars, ablation studies, or complete training hyper-parameters and data-augmentation details, which would improve reproducibility of the reported Dice and topological accuracy numbers.
  2. [Methods] Clarify the exact procedure and any post-processing used for cross-slice linking that converts per-slice segmentations into 3D corallite models; this step is mentioned but not described in sufficient technical detail.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive comments, which highlight important aspects of our evaluation and claims. We address each major comment below, indicating where revisions will be made to improve clarity and acknowledge limitations without overstating the current results.

read point-by-point responses

  1. Referee: [Abstract / Evaluation] Abstract and evaluation sections: the central claim that the method supports 'full 3D individual corallite modelling' rests on 2D per-slice segmentation metrics (Dice 0.77/0.63, topological accuracy 0.94 on held-out slices). No 3D metrics—volumetric IoU, corallite instance tracking precision across slices, or consistency over full colony height—are reported, leaving the cross-slice linking step and its reliability on complete volumes unquantified.

    Authors: We agree that the reported metrics are 2D per-slice segmentation results and that no direct quantitative 3D metrics (such as volumetric IoU or instance tracking across full volumes) are provided. The cross-slice linking is described as a means to produce 3D models from the 2D segmentations, but its reliability on complete colony volumes remains unquantified in the current experiments. This is a genuine limitation of the presented evaluation. In the revised manuscript we will modify the abstract and add a limitations paragraph to clarify that the work demonstrates 2D segmentation as an enabling step for 3D corallite modeling rather than a fully validated end-to-end 3D pipeline, and we will outline suitable 3D metrics for future work. revision: partial

  2. Referee: [Abstract] Abstract: generalization is demonstrated on only two specimens, with a substantial drop to Dice 0.63 on the biologically unrelated specimen. The dataset comprises 697 slices and 37 annotated sections; the authors themselves flag the experiments as limited in scale. Additional colonies and quantitative assessment of 3D reconstruction on full volumes would be required to support the broader claim.

    Authors: We already note in the manuscript that the experiments are limited in scale and context, and the performance drop on the unrelated specimen is reported explicitly. This drop reflects real biological variability between colonies and underscores the challenge of the task. While we concur that additional colonies and full-volume 3D validation would strengthen generalization claims, acquiring and annotating further μCT data from multiple colonies is resource-intensive and lies outside the scope of this initial baseline study. In revision we will expand the discussion of the two specimens' biological differences and the implications of the observed performance gap, but we cannot add new colonies at this stage. revision: partial

standing simulated objections not resolved
  • Additional colonies and full quantitative 3D validation on complete volumes, which would require new data acquisition and annotation beyond current resources.

Circularity Check

0 steps flagged

No circularity detected in CoralLite's empirical pipeline

full rationale

The paper introduces a dataset of μCT slices and a hybrid V-Trans-UNet model for corallite segmentation. It describes pre-training on weakly annotated data, topology-aware fine-tuning on 8k+ manual annotations from 37 slice sections, and evaluation on held-out slices from the same colony plus one unrelated specimen using Dice and topological accuracy metrics. No equations, derivations, or first-principles results are presented that reduce to fitted inputs by construction. There are no self-citation chains, uniqueness theorems, or ansatzes that bear load-bearing claims. The evaluation relies on standard train/test splits with independent test data, making the reported results self-contained without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Central claim rests on the assumption that a hybrid transformer-UNet trained with weak-then-full supervision on a small number of manually annotated slices can produce usable 3D corallite models; no new physical entities or ad-hoc constants are introduced.

axioms (1)
  • domain assumption A hybrid V-Trans-UNet architecture is appropriate for segmenting tiled μCT virtual slabs of Porites sp. colonies
    Invoked when proposing the segmentation model in the abstract.

pith-pipeline@v0.9.0 · 5892 in / 1251 out tokens · 45451 ms · 2026-05-20T20:58:09.086837+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 2 internal anchors

  1. [1]

    Coral Reefs39(4), 1125–1136 (2020)

    Benedetti, M.C., Bramanti, L., Priori, C., Erra, F., Iannelli, M., Bulleri, F., Santan- gelo, G.: Polyp longevity in a precious gorgonian coral: hints toward a demographic approach to polyp dynamics. Coral Reefs39(4), 1125–1136 (2020)

    work page 2020

  2. [2]

    Springer Sci- ence & Business Media (2010)

    Cabioch, G., Davies, P., Done, T., Gischler, E., Macintyre, I., Wood, R., Woodroffe, C.: Encyclopedia of modern coral reefs: structure, form and process. Springer Sci- ence & Business Media (2010)

    work page 2010

  3. [3]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.:Transunet:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021) CoralLite 15

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  4. [4]

    In: International conference on medical image computing and computer-assisted intervention

    Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: International conference on medical image computing and computer-assisted intervention. pp. 424–432. Springer (2016)

    work page 2016

  5. [5]

    Journal of Structural Biology p

    Coronel-Zegarra, A., Knaub, J.L., Merk, V., Pandya, A.: Leveraging deep learning semantic segmentation for imaging coral skeletons. Journal of Structural Biology p. 108313 (2026)

    work page 2026

  6. [6]

    Journal of the Royal Sta- tistical Society

    Cox, D.R.: The regression analysis of binary sequences. Journal of the Royal Sta- tistical Society. Series B (Methodological)20(2), 215–242 (1958)

    work page 1958

  7. [7]

    Marine Biology117(2), 321–326 (1993)

    Darke, W., Barnes, D.: Growth trajectories of corallites and ages of polyps in massive colonies of reef-building corals of the genus porites. Marine Biology117(2), 321–326 (1993)

    work page 1993

  8. [8]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  9. [9]

    Journal of structural biology213(4), 107782 (2021)

    Drake, J.L., Varsano, N., Mass, T.: Genetic basis of stony coral biomineralization: History, trends and future prospects. Journal of structural biology213(4), 107782 (2021)

    work page 2021

  10. [10]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion. pp. 574–584 (2022)

    work page 2022

  11. [11]

    The Holocene13(2), 187–199 (2003)

    Hendy, E., Gagan, M., Lough, J.: Chronological control of coral records using luminescent lines and evidence for non-stationary enso teleconnections in northeast australia. The Holocene13(2), 187–199 (2003)

    work page 2003

  12. [12]

    Frontiers in Insect Science3, 3 (2023)

    Jonsson, T.: Micro-ct and deep learning: Modern techniques and applications in insect morphology and neuroscience. Frontiers in Insect Science3, 3 (2023)

    work page 2023

  13. [13]

    Trends in Ecology & Evolution28(7), 432–441 (2013)

    Kühl, H.S., Burghardt, T.: Animal biometrics: quantifying and detecting pheno- typic appearance. Trends in Ecology & Evolution28(7), 432–441 (2013)

    work page 2013

  14. [14]

    Journal of Experimental Marine Biology and Ecology211(1), 29–67 (1997)

    Lough, J., Barnes, D.: Several centuries of variation in skeletal extension, density and calcification in massive porites colonies from the great barrier reef: a proxy for seawater temperature and a background of variability against which to identify unnatural change. Journal of Experimental Marine Biology and Ecology211(1), 29–67 (1997)

    work page 1997

  15. [15]

    Limnology and Oceanography (2022)

    Medellín-Maldonado, F., López-Pérez, A., Ruiz-Huerta, L., Carricart-Ganivet, J.P.: Understanding corallite demography to comprehend potential bias in scle- rochronology: Analysis of coral modular growth by micro-computed tomography. Limnology and Oceanography (2022)

    work page 2022

  16. [16]

    Milletari, N

    Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neu- ral networks for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV). pp. 565–571 (2016). https://doi.org/10.1109/3DV.2016.79

    work page doi:10.1109/3dv.2016.79 2016

  17. [17]

    Paleoceanography31(10), 1395– 1408 (2016)

    Sadler, J., Webb, G.E., Leonard, N.D., Nothdurft, L.D., Clark, T.R.: Reef core insights into mid-holocene water temperatures of the southern great barrier reef. Paleoceanography31(10), 1395– 1408 (2016). https://doi.org/https://doi.org/10.1002/2016PA002943, https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/2016PA002943

    work page doi:10.1002/2016pa002943 2016

  18. [18]

    arXiv preprint arXiv:2305.15911 (2023) 16 J

    Shi, P., Guo, X., Yang, Y., Ye, C., Ma, T.: Nextou: Efficient topology-aware u-net for medical image segmentation. arXiv preprint arXiv:2305.15911 (2023) 16 J. Jones et al

    work page arXiv 2023

  19. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H.: cldice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16560–16569 (2021)

    work page 2021

  20. [20]

    In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Pro- ceedings, Part III 19

    Würfl,T.,Ghesu,F.C.,Christlein,V.,Maier,A.:Deeplearningcomputedtomogra- phy. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Pro- ceedings, Part III 19. pp. 432–440. Springer (2016)

    work page 2016

  21. [21]

    Coral Reefs41(4), 973–985 (2022)

    Yudelman, E.A., Slowey, N.C.: Coral extension rate analysis using computed axial tomography. Coral Reefs41(4), 973–985 (2022)

    work page 2022

  22. [22]

    In: Fifteenth International Conference on Machine Vision (ICMV 2022)

    Zeng, C., Yang, X., Mirmehdi, M., Gambaruto, A.M., Burghardt, T.: Video- transunet: temporally blended vision transformer for ct vfss instance segmenta- tion. In: Fifteenth International Conference on Machine Vision (ICMV 2022). vol. 12701, pp. 98–105. SPIE (2023)

    work page 2022