CLAIRE-DSA: AI for DSA Quality Control

• CLAIRE-DSA is a deep learning framework that automates classification of fluoroscopic MinIPs from DSA, ensuring quality control in stroke imaging workflows.
• It employs nine dedicated ResNet-based classifiers with weighted cross-entropy to predict image properties like anatomical coverage, projection, and artefact burden.
• When integrated with segmentation models such as CAVE, CLAIRE-DSA filters out low-quality images, boosting segmentation success rates significantly.

CLAIRE-DSA, expanded as Classification AI for Radiological Exams – DSA, is a deep learning–based framework for automated classification of fluoroscopic image properties in minimum intensity projections (MinIPs) derived from digital subtraction angiography (DSA) series acquired during mechanical thrombectomy for acute ischemic stroke. It is designed as a quality-assurance layer for downstream computer-vision pipelines, especially vessel segmentation, by predicting nine image properties that capture anatomical coverage, acquisition geometry, contrast status, and artefact burden. In the reported study, separate classifiers were trained on an annotated dataset of 1,758 fluoroscopic MinIPs from 148 patients, achieving ROC-AUC values from 0.91 to 0.98 across labels; when used to filter unsuitable images before segmentation, the success rate of a downstream CAVE model increased from 42% to 69%, with $p < 0.001$ (Berg et al., 18 Aug 2025).

1. Clinical setting and problem definition

CLAIRE-DSA is situated in the endovascular treatment workflow for acute ischemic stroke. During endovascular thrombectomy, fluoroscopy is used for real-time visualization of catheters and wires, while DSA sequences are acquired for high-resolution visualization of vascular anatomy. In the framework, each DSA series is converted into a single MinIP by taking, for each pixel, the minimum intensity over time; this condenses the series into a representative 2D image that preserves key vascular and anatomical information (Berg et al., 18 Aug 2025).

The framework addresses a practical bottleneck in DSA-based AI. Computer-vision systems for vessel segmentation, occlusion detection, or automated TICI scoring are sensitive to image quality, anatomical coverage, projection angle, contrast visibility, DSA status, and motion artefacts. Poor or unsuitable images can substantially degrade downstream performance, and prior work had to exclude approximately 30% of eligible patients because of inadequate image quality or mismatched properties before automated analysis was possible. CLAIRE-DSA is therefore formulated as an automated image-property classifier whose outputs can be used for filtering, annotation, and workflow optimization rather than as a direct diagnostic model (Berg et al., 18 Aug 2025).

Its operational assumption is narrow and explicit: the input is a fluoroscopic MinIP derived from EVT-related DSA or fluoroscopy, and the output is a set of predictions over nine predefined properties. The system is not described as a general-purpose angiography interpreter, nor as a temporal sequence model over full DSA runs. It functions as a preprocessing and quality-control stage for subsequent algorithms (Berg et al., 18 Aug 2025).

2. Dataset, labels, and annotation protocol

The data originate from the MR CLEAN Registry, a multicenter Dutch EVT registry for acute ischemic stroke covering March 2014 to November 2017. The study used 148 randomly sampled patients and 1,758 fluoroscopic DICOM sequences converted to MinIPs, with intensities normalized to $[0,1]$. The split was stratified at the patient level into training, validation, and test partitions of 70%, 15%, and 15%, respectively, to avoid leakage across patients (Berg et al., 18 Aug 2025).

Each MinIP was annotated for nine labels selected with clinical and imaging experts to reflect properties that strongly affect downstream AI performance.

Label	Classes	Function
Neuro Imaging	Not Neuro; Neuro	Distinguishes neurovascular from non-neuro images
Skull Visibility	Neck; Full; Partial	Encodes cranial coverage
Projection	AP view; Oblique; Left Lat.; Right Lat.	Encodes acquisition geometry
Contrast Fluid	Absent; Present	Encodes vessel opacification
DSA	Not DSA; DSA	Distinguishes subtracted from generic fluoroscopy
Motion Artefact	None; Mild; Severe	Encodes motion degradation
Hemisphere	Indeterm.; Left Hemi; Right Hemi	Encodes side visibility
ICA Top Visible	Not visible; Visible	Encodes a major vascular landmark
MCA Visible	Not visible; Visible	Encodes target vessel visibility

The class distributions are explicitly imbalanced. Examples include 166 Not Neuro versus 1,592 Neuro images, 137 Neck versus 426 Full and 1,029 Partial skull-visibility images, 95 Oblique projections versus 939 AP and 461 Right Lat. images, and 272 Contrast Absent versus 1,321 Contrast Present images. This class imbalance motivated weighted losses during training (Berg et al., 18 Aug 2025).

Annotation quality was evaluated with both multi-rater and expert-comparison statistics. Four raters annotated all 1,758 MinIPs, and an additional experienced rater annotated a randomly sampled subset of 126 images; the subset size was chosen using a minimal sample size formula for Cohen’s $\kappa$ and Fleiss’ $\kappa$, following Donner (1992). Fleiss’ $\kappa$ reached the “reliable” threshold of at least 0.81 for all labels except Skull Visibility ($0.713$), Motion Artefact ($0.594$), and MCA Visible ($0.637$). Cohen’s $\kappa$ between the experienced rater and each other rater reached at least 0.81 for all labels except Motion Artefact ($0.779$) and MCA Visible ($[0,1]$0). This establishes that the most difficult labels for the model are also those with lower annotation consistency (Berg et al., 18 Aug 2025).

3. Architecture, optimization, and implementation

CLAIRE-DSA is not a monolithic multi-label network. It is implemented as nine dedicated ResNet-based classifiers, one per label, each fine-tuned from ImageNet-pretrained backbones. The candidate backbones were ResNet-18, ResNet-34, and ResNet-50, and the model with the lowest validation loss was retained for each label. The selected backbones were ResNet-18 for Neuro Imaging and MCA Visible, and ResNet-34 for Skull Visibility, Projection, Contrast Fluid, DSA, Motion Artefact, Hemisphere, and ICA Top Visible (Berg et al., 18 Aug 2025).

The fine-tuning strategy is deliberately conservative. Only the final three residual blocks and the classification head are trainable; earlier layers remain frozen. This is paired with dropout at rate $[0,1]$1 in the network head, AdamW optimization with weight decay $[0,1]$2, and a fixed training schedule of 10 epochs per model. The paper attributes the choice of 10 epochs to inspection of loss curves and convergence behavior (Berg et al., 18 Aug 2025).

Class imbalance is handled through inverse-frequency weighting in cross-entropy. For binary labels, the loss is binary cross-entropy; for multi-class labels, categorical cross-entropy is used; and the class-weighted form is

$[0,1]$3

with $[0,1]$4, where $[0,1]$5 is the class frequency in the training set (Berg et al., 18 Aug 2025).

The implementation is in Python and was trained on an Nvidia RTX 3060 Ti GPU. The released package includes training and inference scripts for each label, integration examples for applying CLAIRE-DSA to new DSA or fluoroscopic data, and pipeline components for MinIP creation, normalization, and label prediction. Grad-CAM is used for interpretability on selected labels, specifically Skull visibility and Contrast fluid; the reported activation maps indicate attention on skull contours and vessel regions filled with contrast (Berg et al., 18 Aug 2025).

4. Classification behavior and reported performance

Evaluation was performed on the test set using ROC-AUC or macro ROC-AUC, accuracy, precision, recall, and $[0,1]$6-score. Across all nine labels, ROC-AUC ranged from 0.91 to 0.98, and precision ranged from 0.70 to 1.00. The highest-performing labels were those with strong visual cues. DSA versus non-DSA reached ROC-AUC $[0,1]$7, accuracy $[0,1]$8, precision $[0,1]$9, recall $\kappa$0, and $\kappa$1. Contrast Fluid reached ROC-AUC $\kappa$2, accuracy $\kappa$3, precision $\kappa$4, recall $\kappa$5, and $\kappa$6. ICA Top Visible reached ROC-AUC $\kappa$7, accuracy $\kappa$8, precision $\kappa$9, recall $\kappa$0, and $\kappa$1 (Berg et al., 18 Aug 2025).

The more challenging labels align with both lower inter-rater agreement and subtler visual criteria. Projection reached macro ROC-AUC $\kappa$2, but precision and recall were both $\kappa$3, which the paper attributes to difficulty distinguishing minority classes such as oblique views. Motion Artefact reached macro ROC-AUC $\kappa$4, accuracy $\kappa$5, precision $\kappa$6, recall $\kappa$7, and $\kappa$8. MCA Visible reached ROC-AUC $\kappa$9, accuracy $\kappa$0, precision $\kappa$1, recall $\kappa$2, and $\kappa$3. Skull Visibility occupied an intermediate position, with ROC-AUC $\kappa$4, accuracy $\kappa$5, precision $\kappa$6, recall $\kappa$7, and $\kappa$8 (Berg et al., 18 Aug 2025).

The paper explicitly characterizes the easiest labels as DSA versus non-DSA, Contrast fluid, ICA top visibility, and Hemisphere, and the hardest as Projection, Motion artefact severity, and MCA visibility. This suggests that CLAIRE-DSA is strongest when the target property is associated with stable global structure or a strong radiographic cue, and weaker when the target depends on subjective gradation or small-vessel visibility (Berg et al., 18 Aug 2025).

5. Downstream segmentation, filtering logic, and workflow role

The practical utility of CLAIRE-DSA was evaluated using the Cerebral Artery-Vein Segmentation (CAVE) model as a downstream task. On the same external test set of 234 MinIPs, each image was processed both by CAVE and by CLAIRE-DSA, and an experienced rater assessed whether the segmentation was successful. A segmentation was considered successful if visual inspection indicated that at most 20% of pixels were incorrectly classified (Berg et al., 18 Aug 2025).

CLAIRE-DSA was then used as a pre-filter for segmentation suitability. An image was deemed unsuitable if any of the following predicted conditions held: Neuro Imaging = Not Neuro; Skull Visibility = Neck; Contrast Fluid = Absent; DSA = Not DSA; or Motion Artefact = Mild or Severe. All remaining images were considered suitable. Under this rule set, the unfiltered test set had a segmentation success rate of 42%, whereas the CLAIRE-DSA-filtered subset contained 87 images and had a success rate of 69%, with a Z-test for proportions yielding $\kappa$9 (Berg et al., 18 Aug 2025).

The reported confusion matrix further clarifies the operating characteristics of the filter. CLAIRE-DSA marked 60 images as usable that CAVE segmented correctly, 27 as usable that CAVE segmented incorrectly, 37 as unusable that CAVE would actually have segmented correctly, and 110 as unusable that CAVE segmented incorrectly. The derived sensitivity was approximately $0.713$0, while specificity was approximately $0.713$1. The paper interprets this as CLAIRE-DSA correctly filtering out 80% of images that would lead to failed segmentation, while missing about 38% of images that could have been segmented correctly. This suggests a conservative filter oriented toward excluding problematic inputs rather than maximizing retention (Berg et al., 18 Aug 2025).

The framework is therefore best understood as an enabling component for data curation, automatic annotation, and procedural quality control. The authors discuss immediate feedback during acquisition, guidance for re-acquisition, multicenter quality standardization, and preprocessing for research datasets. They also note limitations: the dataset contains 1,758 MinIPs from 148 patients; all data come from a single registry in the Netherlands; some labels remain subjective; and the current filtering rules were not optimized for any single downstream model. Future directions include larger datasets, sequence-based modeling, joint quality-and-segmentation models, adaptation to other modalities or interventional procedures, and real-time deployment with PACS or interventional-suite consoles (Berg et al., 18 Aug 2025).

6. Terminology, scope, and related uses of the name

Within the arXiv literature, the exact name CLAIRE-DSA is not stable across domains. The formal expansion Classification AI for Radiological Exams – DSA is specific to the 2025 fluoroscopic-image classification framework for acute ischemic stroke (Berg et al., 18 Aug 2025). By contrast, the earlier paper “The Cross-Lingual Arabic Information REtrieval (CLAIRE) System” defines CLAIRE as an end-to-end cross-lingual Arabic information-retrieval system, and the accompanying technical description states explicitly that the string “CLAIRE-DSA” does not appear anywhere in that paper, that there is no module or variant named “CLAIRE-DSA,” and that “DSA” is not defined there (Chen et al., 2021).

A similar caveat applies in the French-language-model literature. “The Claire French Dialogue Dataset” presents CFDD as the dialogue-focused French corpus used by the Claire model family, and the description notes that “CLAIRE-DSA” as a named entity does not occur; the central artifact is CFDD rather than a separate CLAIRE-DSA component (Hunter et al., 2023).

Later DSA-focused medical-imaging papers complicate the name further. “Semantically Conditioned Diffusion Models for Cerebral DSA Synthesis” is described as the basis for what one could call a CLAIRE-DSA system, but the paper itself does not explicitly use the acronym (Xu et al., 12 Feb 2026). “Leptomeningeal Collateral Detection on DSA via Vessel-Graph Neural Networks” likewise presents a complete DSA analysis pipeline and explicitly notes that the authors do not introduce the name “CLAIRE-DSA” (Cao et al., 12 Jun 2026). Beyond medical imaging, DSA is also an established acronym in unrelated literatures, including decentralized stochastic optimization (Mokhtari et al., 2015), digital signatures (Zahhafi et al., 2021), and diffusion synthetic acceleration for transport solvers (Haut et al., 2018). The most precise contemporary usage of CLAIRE-DSA therefore refers to the 2025 ResNet-based MinIP quality-classification framework for EVT-related DSA in acute ischemic stroke, rather than to the broader CLAIRE family or to other senses of DSA.