ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography
summary
The gist
Accurate segmentation of the coronary vessel lumen is a prerequisite for quantitative assessment of atherosclerotic plaque and perivascular adipose tissue in coronary computed tomography angiography…
This episode discusses
- ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography · Paper Radio
The paper
ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography · Read on arXiv
Kit M. Bransby, Esther Øksnebjerg, Kristoffer Kjær, Jacob Kirkeby, Yasmin El Youssef, Aïda Jiménez, Philip R. Pedersson, Martina C. de Knegt, Klaus F. Kofoed
Technical University of Denmark, DTU Compute, Kongens Lyngby, Denmark · Copenhagen University Hospital – Rigshospitalet, Copenhagen, Denmark
Accurate segmentation of the coronary vessel lumen is a prerequisite for quantitative assessment of atherosclerotic plaque and perivascular adipose tissue in coronary computed tomography angiography (CCTA). Cardiologists rely on semi-automated methods for this task because manual vessel tracing and segmentation are labour-intensive. Although many automated methods have been proposed, their validation remains limited by the lack of large, high-quality publicly available datasets. We provide a new dataset of voxel-wise annotations of the vessel lumen and coronary segments, alongside centerlines, and mesh surfaces for 800 scans from the publicly available ImageCAS dataset. Using this dataset, we benchmark established lumen segmentation methods against inter-observer variability, stratifying performance by disease, image quality, coronary dominance, coronary segment, vessel diameter, and lumen attenuation. These labels allow segmentation accuracy to be described in anatomical and clinical context rather than reported as a single aggregate score. The dataset supports the development and validation of methods for lumen segmentation, plaque and perivascular quantification, and haemodynamic modelling.
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography".
Jane: Here's the gist of the paper. The paper details a comprehensive shared training and inference framework designed for coronary artery segmentation and centerline extraction, ensuring that "all methods are trained and evaluated through a single shared framework that fixes everything which must remain identical for a fair comparison." Data Input and Preprocessing: The input data involves processing the cMPR volume.
Tom: Specifically, "Random crops from cMPR volume of size N × H × W were used as input," where N was set at 10 frames with 0.4mm spacing between frames, and H and W were set at 64, representing a "16 × 16mm cross section using 0.25 mm pixel spacing." For the shared framework (Section E), every scan is first processed by resampling to "isotropic 0.5 mm spacing," clipping intensities to a "
−200, 1000: HU window and linearly mapped to
0, 1: ." The framework supports three input types: "whole volume where the full preprocessed volume is passed to the network; random crop where a fixed-size patch is cropped at a random location from the whole volume; and center crop where patches are extracted at precomputed coordinates." For patch-based methods, "foreground (fg) sampling is used to ensure at least 50% of every batch contains patches with lumen class." Anatomical Segmentation Classes: The study utilizes a defined set of coronary segments, which are mapped to specific class indices: Branch Segment Class ID:---:---:--- Left main LM (Left main) 1 Left anterior descending artery LAD (LAD) 2-3 (Implied range/separate classes) *
Note: The table lists LAD as 2: * Left circumflex artery LCx (LCx) 3-4 *
Note: The table lists LCx as 3: * First diagonal branch D1 (D1) 4-5 *
Note: The table lists D1 as 4: * Second diagonal branch D2 (D2) 5-6 *
Note: The table lists D2 as 5: * First obtuse marginal branch OM1 (OM1) 6-7 *
Note: The table lists OM1 as 6: * Second obtuse marginal branch OM2 (OM2) 7-8 *
Note: The table lists OM2 as 7: * Intermediate marginal (ramus intermedius) IM (IM) 8-9 *
Note: The table lists IM as 8: * Right coronary artery RCA (RCA) 9-10 *
Note: The table lists RCA as 9: * Right posterior descending artery R-PDA (R-PDA) 10-11 *
Note: The table lists R-PDA as 10: * Right posterior lateral artery R-PLA (R-PLA) 11-*
Note: The table lists R-PLA as 12: * Left posterior descending artery L-PDA (L-PDA) 12-*
Note: The table lists L-PDA as 13: * Left posterior lateral artery L-PLA (L-PLA) 13-*
Note: The table lists L-PLA as 14: * Other - - Training and Augmentation Pipeline: Augmentation is applied during training based on the nnU-Net’s default pipeline, including: "additive Gaussian noise (p=0.1, sigma in
0, 0.01: ), Gaussian blur (p=0.2, sigma in
0.5, 1.0: ), multiplicative brightness (p=0.15,
0.75, 1.25: ), contrast (p=0.15,
0.75, 1.25: ), simulated low resolution (p=0.25, downsample factor
0.5, 1.0: ), inverted gamma (p=0.1, gamma in
0.7, 1.5: ) and gamma (p=0.3, gamma in
0.7, 1.5: )." Notably, "Axis flipping over coronal and sagittal planes are included," while rotations and axial flipping are omitted because "coronary anatomy has a genuine superior–inferior orientation." The training schedule defines an epoch as "250 training and 50 validation batches drawn with replacement." Most methods use Adam (with weight decay 10-8) and a polynomial learning-rate decay (power 0.9), while nnU-Net used SGD.
Jane: Inference Pipeline: For inference, the process varies by method type: "whole-volume methods used a single forward pass, while patch- and crop-based methods are tiled with a sliding window of the method’s own training patch size at 50% overlap." Overlapping predictions are combined using a "Gaussian-weighted average with per-axis sigma = 0.125 times patch size." A critical step is "Axis mirroring test-time augmentation," where "predictions are averaged over all four flip combinations of the two in-plane axes, matching the flip axes used during training." Finally, post-processing is applied at full original resolution and remains identical for all methods: a "threshold at 0.5, followed by removal of connected components smaller than 100 voxels." Method-Specific Implementation Notes: The framework required
Paper discussion segment 2: Jane: To pick up where we left off, after discussing the sheer value of ImageCAS-X as a resource—a benchmark, essentially—we need to look at what the authors are telling us about *how* we should be using this data.
Tom: The core message from reviewing the paper’s summary is that this dataset isn't just a passive collection of images; it’s an active tool that forces researchers to refine their methodology and prove reliability in ways they hadn't before.
Jane: It emphasizes that achieving true clinical reliability requires more than just feeding raw data into the most powerful general AI model available today. The research community needs to shift its focus toward validating algorithms under highly constrained, medically plausible scenarios.
Lu: What I found particularly insightful in the summary is how it highlights the need for quantitative metrics that go beyond simple pixel accuracy. We need to measure structural continuity and functional plausibility—the algorithm must respect the physics of blood flow, not just statistical likelihood.
Meng: And this speaks directly to robustness in implementation. The summary implies that if an algorithm fails in a specific topological area, it shouldn't fail completely; it needs targeted components that can isolate and correctly handle those complex junctions or delicate side branches gracefully.
Lalam: From a resource utilization standpoint, the authors are implicitly stating that this dataset will accelerate the development of specialized diagnostic tools worldwide. Instead of each hospital needing to build its own validation set, they have a shared, reliable starting point.
Tom: So, if I understand correctly, the summary is moving us from asking "Can AI segment this?" to asking "Does AI segment this *physiologically* accurately?"
Jane: Precisely. The authors are setting a higher bar for proof-of-concept than ever before. They are guiding researchers toward specialized, verifiable solutions rather than general, massive models that might perform well on average but fail catastrophically in edge cases.
Tom: This brings us to the next logical question: what specific improvements does this benchmark force us to adopt? Let’s dive into the technical suggestions from the paper.
Paper discussion segment 3: Tom: We've established that ImageCAS-X provides a fantastic resource and that its summary points toward a methodological shift; now we need to talk about what concrete improvements the authors are demanding of us.
Jane: The key takeaway here is that the next generation of coronary analysis tools cannot be monolithic algorithms trying to solve every problem simultaneously. They must be designed with specialized, modular components built specifically for known anatomical difficulties.
Lu: This concept of modularity is critical because human anatomy is inherently irregular and redundant. Instead of treating the entire coronary tree as one continuous problem, the system needs specialized modules—one for major trunks, one for complex junctions, and another dedicated just to those tiny side branches.
Meng: From an engineering perspective, this architecture significantly improves error handling. If the module responsible for tracing a sharp bend encounters noise or incomplete data, it can flag that specific topological weakness without crashing the entire analysis pipeline.
Lalam: Considering global implementation, a modular approach is also more adaptable to different hardware and software environments found in various medical facilities worldwide. You can upgrade one specialized component without having to overhaul the entire system.
Tom: It’s about building constraints into the model itself—making sure that the AI doesn't just produce a plausible-looking image, but an anatomically *possible* one.
Jane: Exactly. The paper is forcing researchers to integrate core biomechanical knowledge directly into the computational framework. The algorithm needs to know, for example, that vessel curvature must adhere to certain physical limits and that the flow path must remain continuous and physically sound.
Lu: That requirement for continuous physical plausibility fundamentally changes how we define success. It's moving us beyond just counting pixels correctly toward validating a physically coherent model of blood flow itself.
Meng: This elevates the design requirements from simply being mathematically accurate to being biomechanically accurate, which is a massive leap in engineering complexity and reliability.
Tom: So, we are building systems that don't just read data; they *understand* human biology. And this leads us perfectly into the question of how these technically superior tools will translate into actual clinical workflows.
Paper discussion segment 3: Tom: We’ve been discussing the
Conclusion: Tom: So, after diving deep into the methodology and findings today, it’s clear that this paper has set a new standard for how we approach coronary artery analysis moving forward.
Jane: Exactly. Ultimately, "ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography" provides the necessary rigorous framework that the entire clinical community needed to move forward with confidence.
Lu: For me, the implications are massive because mapping out that continuous, branching topology genuinely opens up computational possibilities we hadn't even considered before.
Meng: From an engineering viewpoint, this gives us a truly stable and reliable foundation—a massive step toward actual deployment in clinical settings where guaranteed accuracy is essential.
Lalam: And from a global perspective, it’s more than just data; it establishes a shared knowledge base that elevates the entire field of cardiovascular imaging worldwide.
Tom: It really does force us to validate automated methods against established human agreement in the most complex scenarios, which is non-negotiable for patient safety.
Jane: Knowing precisely where those AI models struggle is as valuable as knowing where they succeed; it provides a direct roadmap for future development and refinement.
Lu: I genuinely cannot wait to see researchers build better systems using this level of structural detail; I think it opens up avenues we haven't even considered yet in terms of functional modeling.
Meng: We are seeing a clear path toward deployment that is far more robust and reliable than the manual methods we previously relied on, which is an enormous win for efficiency across healthcare systems.
Lalam: This signifies a major shift in how medical knowledge is shared and utilized globally—a massive win for advancing healthcare across all cultures.
Tom: I think we can all agree this foundational work will redefine our standards for accuracy in coronary artery analysis moving forward, giving us unprecedented confidence.
Jane: It provides the standardized reference point that the clinical community needed, allowing them to proceed with much greater rigor in their diagnoses and research.
Lalam: This structural focus really allows us to improve our fundamental understanding of human biology at a microscopic level, marking a huge societal shift in care.
Tom: Well, team, thank you all for such an insightful deep dive into "ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography."
Jane: It’s been a fantastic discussion that truly underscores the maturity and potential of this field.
Tom: We’ll be moving on to our next paper now, but I hope you all have a great break and enjoy continuing your deep dive into cardiovascular imaging with us!
More episodes
- 2603.28113-The multiplicative Lipschitz bound as an inductive bias
- 2603.28595-Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes
- 2603.29765-Training-Free Dynamic Upcycling of Expert Language Models
- 2605.00778-Observable Performance Does Not Fully Reflect Adaptive System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
- 2605.06906-TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
- 2605.09165-Sparse Layers are Critical to Scaling Looped Language Models
- 2605.09439-Inverse Design for Conditional Distribution Matching
- 2605.10313-Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards: Regret Bound and Statistical Inference
- 2605.10878-Minimum Neural Weight Norm = Kolmogorov Complexity
- 2605.11617-MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound