💻 computer science

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

This paper introduces CoralLite, a novel dataset and deep learning baseline utilizing a hybrid V-Trans-UNet architecture to achieve the first successful 3D reconstruction of individual coral corallites from {\mu}CT scans of entire colonies, thereby enabling the tracking of polyp division and skeletal growth history.

Original authors: Jess Jones, Leonardo Bertini, Kenneth Johnson, Erica Hendy, Tilo Burghardt

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jess Jones, Leonardo Bertini, Kenneth Johnson, Erica Hendy, Tilo Burghardt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a coral colony as a massive, ancient apartment building made of stone. This building isn't just one big block; it's built by thousands of tiny, individual tenants (called polyps) that have been dividing and moving in for hundreds of years. Each tenant lives in their own little stone tube, known as a corallite.

The problem is that this "apartment building" is buried inside a solid rock. Scientists want to study the history of the building by looking at how these individual tubes are arranged, how they grew, and how the tenants divided over time. But looking at the whole rock is like trying to read a book while it's still wrapped in a thick, opaque blanket.

To see inside, scientists use a super-powerful X-ray scanner called micro-CT (µCT). This scanner takes thousands of thin, 2D "slices" of the rock, like slicing a loaf of bread. The result is a giant stack of images showing the tiny stone tubes.

The Challenge: The "Needle in a Haystack" Problem
Manually tracing every single tube in these thousands of slices is like trying to find a specific needle in a haystack, but the haystack is the size of a mountain, and the needles are moving. It takes humans years to do this. The researchers wanted to teach a computer to do it automatically.

The Solution: CoralLite
The paper introduces CoralLite, a new tool and a dataset designed to teach computers how to find and trace these coral tubes automatically. Here is how they did it, using some simple analogies:

1. The Training Data (The "Textbook")

You can't teach a student without a textbook. The researchers created a massive dataset:

The Raw Material: They took 697 high-resolution X-ray slices of a real coral colony.
The Annotations: They manually drew the outlines of over 8,000 individual coral tubes on some of these slices. This is the "answer key" the computer needs to learn from.
The Strategy: They didn't just show the computer the whole rock at once. They chopped the images into small, manageable "tiles" (like puzzle pieces) so the computer could focus on one neighborhood at a time.

2. The Brain (The "Hybrid Architect")

To solve this puzzle, they built a special type of AI brain called a V-Trans-UNet. Think of this brain as having two superpowers working together:

The Local Detective (ResNet): This part looks at the tiny details in a single slice, like the texture of the stone.
The Contextual Reader (Transformer): This part looks at the "big picture" and remembers what it saw in the slice before and the slice after. Just like how you understand a sentence better if you read the words before and after it, this AI understands the 3D shape of a tube by looking at the slices above and below it.

3. The "Glue" Problem (Topology Loss)

Here is the tricky part: In a coral colony, tubes are packed tightly together. A normal AI might get confused and think two neighboring tubes are actually one big, merged blob.

The Fix: The researchers added a special rule (a "topological loss") to the training. Imagine a strict teacher who says, "If you draw two houses touching each other, you get a penalty!" This forces the AI to keep the boundaries of each tube distinct, ensuring it doesn't accidentally glue neighbors together.

4. The Results (The "Magic Reveal")

Once trained, the AI was tested on slices it had never seen before.

Same Building, Different Angle: When tested on the same coral colony but from a different angle, it was very accurate (about 94% topological accuracy).
Different Building: When tested on a completely different species of coral, it still did a decent job (about 63% accuracy), proving it learned the general concept of "coral tubes," not just memorized one specific rock.
The 3D Model: Finally, the computer stitched these 2D slices back together. It successfully reconstructed a 3D model of the entire coral colony, showing thousands of individual tubes as if they were separate, glowing pipes.

What This Means (And What It Doesn't)

The paper claims this is the first time visual machine learning has successfully built a full 3D model of individual coral tubes from X-ray scans alone.

What they achieved: They proved it's possible to automate the "slicing and tracing" process. They created a baseline (a starting point) for the scientific community to build upon.
What they didn't claim: They are not saying this system is perfect yet. They admit that for different species of coral, the accuracy drops. They also note that while the 3D models look great for visual inspection (letting scientists see the structure), they haven't yet used this to calculate exact biological statistics (like exactly how fast the coral is growing) because the dataset is still too small for that level of precision.

In short, CoralLite is a proof-of-concept that shows computers can learn to "see" inside a coral rock and map out its tiny, ancient apartments, saving scientists years of manual work and opening the door to understanding how these vital ecosystems grow and change.

Technical Summary: CoralLite

Problem Statement
The life history of reef-forming corals, such as Porites species, is archived within their accreting skeletons. Understanding the rate and timing of polyp division and its impact on colony growth requires tracking the skeletal corallite deposited around each individual polyp. While these colonies can grow for hundreds of years, their living tissue is a thin surface layer. Currently, analyzing the internal 3D structure of these colonies via micro-computed tomography (µCT) relies on manual tracing of individual corallites through volumetric slices. This process is time-consuming, laborious, and impractical for large-scale analysis. There is a lack of automated, deep learning-based methods capable of performing identity-preserving 3D reconstruction of individual corallites from µCT data alone.

Methodology
The authors propose CoralLite, a pipeline comprising a novel dataset and a hybrid deep learning architecture for volumetric segmentation and 3D reconstruction.

Dataset Construction:
- The dataset includes 697 µCT slices from three Porites specimens (Naturalis 6785, Naturalis 6781, and an Astraeiformis specimen).
- Annotations include 8,412 manual binary segmentations of corallite interiors across full slices and point-annotation priors (corallite centers) for 33 slices.
- Data is processed into $5 \times 224 \times 224$ tiled volumetric snippets to capture local 3D context without downsampling.
Architecture (Hybrid V-Trans-UNet):
- The model adapts the Video-TransUNet architecture, reinterpreting the temporal dimension as spatial depth.
- Backbone: A ResNet-50 encoder extracts per-slice features.
- Context Module: A Temporal Context Module (TCM) blends features across the depth dimension (adjacent slices) to capture local volumetric continuity.
- Transformer: A 12-layer Vision Transformer (ViT) models longer-range spatial dependencies.
- Decoder: A cascaded deconvolutional DCNN with skip connections produces the final segmentation mask.
Training Strategy:
- Two-Stage Optimization: The model is first pre-trained on weakly annotated data (Naturalis 6785) using standard Binary Cross-Entropy and Dice losses. It is then fine-tuned on a single fully annotated slice from the target colony (Naturalis 6781 Growth #1279).
- Topology-Aware Loss: To address the specific requirement that neighboring corallites must remain separated, a topological loss ( $L_{Topo}$ ) is introduced during fine-tuning. This loss penalizes merged regions and over-expanded false positives by analyzing connected components and their intersections with ground truth.
- Data Augmentation: Includes random rotations, flips, and additive Gaussian noise.
3D Reconstruction:
- A cross-slice linking procedure connects segmented regions across adjacent slices based on centroid distance and Intersection-over-Union (IoU) thresholds.
- Traced regions are represented by centroids and axis lengths to generate 3D elliptical models of individual corallites, visualized as full colony volumes.

Key Contributions

CoralLite Dataset: The first annotated µCT dataset for corallite reconstruction, containing 697 slices, 37 annotated slices (full and partial), and 8,412 manual corallite region labels.
Topology-Guided Pipeline: A volumetric segmentation pipeline utilizing a hybrid ResNet-ViT architecture with a specific topological loss to prevent the merging of adjacent corallite instances.
Baseline Evaluation: Comprehensive evaluation across three generalization settings: same colony/same axis, same colony/different axis, and cross-specimen.
Open Source: Release of all network weights, source code, and the full dataset to serve as a baseline for future research.

Results

Segmentation Performance: The fine-tuned model achieves a mean Dice Similarity Coefficient (mDSC) of 0.77 on the same colony (Task A) and 0.71 on a different projection axis of the same colony (Task B). It achieves 0.63 mDSC on a biologically unrelated specimen (Task C).
Topological Accuracy: The model reaches a topological accuracy of 0.94 on the same colony.
Impact of Components:
- Fine-tuning on a single fully annotated slice is critical; pre-trained models without fine-tuning perform poorly.
- The topological loss does not significantly alter mDSC scores but visibly improves the separation of neighboring corallites, reducing false positive connections.
- Gaussian noise augmentation improves cross-species performance (Task C) by up to 8%.
Reconstruction: The pipeline successfully generates 3D visualizations of individual corallites and colony-scale models, capturing growth directions and structural alignment, though these are currently for visual inspection rather than scientific quantification.

Significance and Claims
The paper presents CoralLite as a proof-of-concept demonstrating that visual machine learning can effectively support full 3D individual corallite modeling from µCT scans alone. The authors modestly claim that their work establishes the first baseline for this task, showing that a reasonably sized annotated dataset combined with volumetric learning is sufficient to support automated colony reconstruction at the corallite level.

The authors explicitly state that they do not yet claim to provide a fully automated, scientifically quantified imageomics application capable of reliably estimating complex biological indicators (such as budding rate variability or polyp lifespan) as a function of colony age. Instead, the paper aims to provide the foundational datasets, annotations, and baselines necessary for the community to build toward such applications. The work marks a first step toward imageomic automation to better understand coral growth dynamics and support reef conservation efforts.