Source-linked AI summary
Automated Distinction of Intimal and Medial Intracranial Arterial Calcification from CT Head
Benjamin Jin, Maria del C. Valdés Hernández, Richard Bortsov, Joanna M. Wardlaw, Daniel Bos, Grant Mair
TL;DR
IAC is associated with neurovascular disease, but prior assessments have chiefly focused on presence and volume rather than subtype. This paper compares three automated classification approaches from segmentation masks and reports broadly similar performance, with segmentation-origin differences in weighted F1 not significant.
Problem
IAC is associated with neurovascular disease, but prior associations have primarily used presence and volume despite evidence that subtype shapes may reflect distinct processes and consequences.
Source: §1 · p. 1Method
The study compares visual-score-inspired, sphericity-based, and medical shape foundation-model embedding approaches for classifying IAC subtypes from segmentation masks.
Source: §1 · p. 2, §2.3 · p. 4Results
The methods achieved relatively similar classification results, and weighted F1 differences between manual and automated segmentation masks were not significant.
Source: §3.1 · p. 5, §3.2 · p. 8Takeaways & Limitations
Combining subtype classification with automated IAC segmentation supports fully automated CT-based quantification, with potential for large-scale clinical datasets.
Source: §3.3 · p. 8, §4 · p. 9The Sphericity method remains bound to the heuristic that intimal calcifications appear more spherical than medial calcifications.
Source: §3.3 · p. 8
Abstract
from arXiv · showhide
Intracranial arterial calcifications (IACs) are a common finding on clinical non-contrast enhanced head CT scans and are associated with neurovascular disease. Calcifications can occur in the intimal or medial layer of the arterial wall, subtypes that differ in aetiology and may have distinct clinical relevance. These subtypes can be visually distinguished by radiologists based on the shape of the calcifications. We investigate three automated approaches for subtype classification of IAC from head CT-derived segmentation masks: (1) an automated adaptation of the established radiological visual score, (2) a sphericity-based method, and (3) a method based on shape embeddings extracted by a medical shape foundation model. All approaches use the same lightweight classification pipeline on top of the features they compute and are evaluated using 5-fold cross-validation. The three methods achieved comparable performance, with the embedding-based approach yielding the best overall results with a weighted F1 (mean $\pm$ SD) of up to 71.5 $\pm$ 3.7 for a single artery and 59.8 $\pm$ 1.7 for the joint artery classification. Performance was largely preserved when using automated instead of manual IAC segmentation masks, and we found the difference in weighted F1 not significant. Our results show that fully automated IAC subtype quantification from head CT is feasible and remains robust to the use of manual and automated IAC segmentation masks. Code at https://github.com/bjin96/iac-subtyping.
1 Introduction
Intracranial arterial calcification (IAC) is visible on routine head CT and associated with neurovascular disease, but associations have chiefly relied on IAC presence and volume. This paper proposes automated classification of intimal and medial IAC from segmentation masks, building on radiological visual assessment.
1 Introduction
- IAC is visible on routine clinical head CT and associated with neurovascular diseases, while prior associations have primarily used its presence and volume.Source: §1 · p. 1
- Figure 1 presents CT examples of IAC subtypes from the cavernous internal carotid artery.Source: Figure 1 · p. 2
- Different calcification shapes may reflect distinct processes with divergent downstream consequences; intimal and medial IAC differ in arterial-wall location and aetiology.Intimal calcification occurs alongside fatty and fibrous material in the intimal layer; medial calcification involves the internal elastic lamina and medial layer.Source: §1 · p. 2
- The paper proposes automated approaches from IAC segmentation masks, extending a previously developed visual score to distinguish intimal and medial calcification.The authors describe this as the first study investigating automated classification of intimal and medial IAC from head CT.Source: §1 · p. 2
2 Methods
The methods derive IAC subtype features from segmentation masks and compare three feature approaches using a shared classification pipeline. Data include visually assessed subtype labels and manual and automated segmentations.
2 Methods
- The methods assume IAC segmentation masks and generate features to distinguish IAC subtypes.Source: §2 · p. 2
2.1 Data
- The Rotterdam Study provided 2,524 participants with non-contrast CT scans from 2003–2006 and visual subtype labels by artery and at patient level, including a mixed category.The mixed category indicates that each subtype was assigned to one of the individual internal carotid arteries; cases with calcification under 4 mm3 were excluded because they can destabilize analyses, especially shape embeddings.Source: §2.1 · pp. 2–3
2.2 Automated Visual Method
- The Automated Visual method computes scan-level features for circularity, thickness, and morphology from skeletonized calcification paths.Morphology uses longest-path length as a proxy for continuity; object-level features are averaged across objects within each CT image series.Source: §2.2 · pp. 3–4
2.3 Sphericity Method
- The Sphericity method compares each calcification object's surface area with that of a volume-matched sphere, then contrasts scan-level distributions with subtype reference distributions using a Wasserstein difference.The resulting Wasserstein difference is used as a classification feature.Source: §2.3 · p. 4
2.4 Deep Learning Method
- The Deep Learning method uses MedShapeNet Point Cloud Encoder embeddings from point clouds derived from IAC masks as shape features.The embedding dimension is reduced from 4096 to 128.Source: §2.4 · p. 4
2.5 Classification Pipeline
- All approaches use logistic regression with 5-fold cross-validation, evaluated separately for left, right, and joint artery classification.Table 1 reports macro- and weighted-average metrics and weighted F1 for automatically generated segmentations.Source: §2.5 · p. 5, Table 1 · p. 5
3 Results
The three approaches produced relatively similar classification results, while pairwise agreement differed; weighted F1 differences between manual and automated segmentation masks were not significant. The discussion describes the approaches’ trade-offs and their compatibility with an automated pipeline.
3.1 Intimal and Medial Calcification Classification
- Mean sphericity correlated with circularity (ρ = −0.62) and morphology (ρ = −0.70), but not thickness (ρ = 0.00).The authors suggest the lower agreement of Deep Learning with the two hand-crafted methods may reflect complementary shape characteristics.Source: §3.1 · p. 6, Figure 3 · p. 6
- Automated Visual tended to confuse medial with intimal IAC, whereas Sphericity and Deep Learning balanced performance between the subtypes.All three methods struggled to identify the mixed class, which denotes different predominant subtypes in individual internal carotid arteries.Source: §3.1 · pp. 6–7, Figure 4 · p. 7
- Classification results were relatively similar across methods, but Automated Visual and Sphericity agreed most, with both agreeing less with Deep Learning.Table 2 reports pairwise agreement; the passage does not provide its numerical values.Source: §3.1 · p. 5, Table 2 · p. 6
3.2 Manual and Automated Segmentation Masks
- Manual and automated masks produced slightly different mean-sphericity distributions, but retained contrasting subtype shapes; weighted F1 differences were not significant.The comparison used bootstrap resampling with 10,000 samples and a 95% confidence interval for mean differences.Source: Figure 5 · p. 7, §3.2 · p. 8
3.3 Discussion and Limitations
- Manual visual assessment is practical for small to medium datasets, but its coarse artery-level scoring and scaling manual labour can make population-level use costly.Related morphometric methods rely on CT angiography and lack subtype information.Source: §3.3 · p. 8
- Sphericity remains tied to the heuristic that intimal calcifications are more spherical, while shape embeddings offer flexibility at the cost of explainability.The authors note that the point-cloud encoder could be adapted for tiny objects, or those objects could be handled separately in future classification pipelines.Source: §3.3 · p. 8
- The approach is compatible with automated segmentation, supporting a fully automated pipeline for IAC subtype quantification from head CT.The authors state that fully automated description enables large-scale subtype quantification and downstream association with neurovascular pathology and clinical outcomes.Source: §3.3 · p. 8
4 Conclusion
The study proposes three automated approaches for distinguishing intimal and medial IAC. Combined with automated segmentation, they support subtype quantification directly from CT scans.
4 Conclusion
- The study proposes three approaches to automating the distinction between intimal and medial IAC.Source: §4 · p. 8
- Combining the approaches with automated IAC segmentation supports fully automated subtype quantification directly from CT scans.Source: §4 · pp. 8–9
- Automated subtype quantification could support population-level evidence on IAC and neurovascular disease or recruitment of at-risk participants for treatment research.Source: §4 · p. 9
- In the longer term, automated IAC subtype quantification could inform individual patient stratification in clinical practice.Source: §4 · p. 9