CLAIRE-DSA: Fluoroscopic Image Classification for Quality Assurance of Computer Vision Pipelines in Acute Ischemic Stroke

TL;DR

The paper tackles the problem of variable fluoroscopic image quality in DSA during acute ischemic stroke treatment and its impact on AI pipelines. It introduces CLAIRE-DSA, a multi-label classification framework using nine ResNet-based backbones fine-tuned from ImageNet to predict image properties from fluoroscopic MinIPs. On a dataset of $1{,}758$ MinIPs, CLAIRE-DSA achieves ROC-AUC values between $0.91$ and $0.98$ across labels and improves a downstream cerebral vessel segmentation task from $42\%$ to $69\%$ success ($p<0.001$) by filtering out unsuitable images. The work demonstrates the practical utility of automated image quality control in DSA-based stroke care and provides open-source code to facilitate adoption in clinical and research settings.

Abstract

Computer vision models can be used to assist during mechanical thrombectomy (MT) for acute ischemic stroke (AIS), but poor image quality often degrades performance. This work presents CLAIRE-DSA, a deep learning--based framework designed to categorize key image properties in minimum intensity projections (MinIPs) acquired during MT for AIS, supporting downstream quality control and workflow optimization. CLAIRE-DSA uses pre-trained ResNet backbone models, fine-tuned to predict nine image properties (e.g., presence of contrast, projection angle, motion artefact severity). Separate classifiers were trained on an annotated dataset containing $1,758$ fluoroscopic MinIPs. The model achieved excellent performance on all labels, with ROC-AUC ranging from $0.91$ to $0.98$, and precision ranging from $0.70$ to $1.00$. The ability of CLAIRE-DSA to identify suitable images was evaluated on a segmentation task by filtering poor quality images and comparing segmentation performance on filtered and unfiltered datasets. Segmentation success rate increased from $42%$ to $69%$, $p < 0.001$. CLAIRE-DSA demonstrates strong potential as an automated tool for accurately classifying image properties in DSA series of acute ischemic stroke patients, supporting image annotation and quality control in clinical and research applications. Source code is available at https://gitlab.com/icai-stroke-lab/wp3_neurointerventional_ai/claire-dsa.

Figures (4)

• Figure 1: Schematic overview of the model pipeline, illustrating data input and pre-processing, backbone architecture and model training and evaluation. First, MinIPs are created, which are then labeled by four raters and split into train, test and validation sets. For each of nine DSA labels, three pretrained ResNet models are finetuned, and the best performing model is selected. CLAIRE-DSA combines the 9 best models.
• Figure 2: Selected Grad-CAM heatmaps for the Skull Visibility (left) and Contrast Fluid (right) labels, showing model attention regions based on the gradient output. For Skull Visibility, the model clearly looks at the edges of the image to see whether the entire skull is present. For Contrast Fluid, the model attends to contrast, as expected.
• Figure 3: Selected ROC curves for the detection of Contrast fluid and Projection. For contrast, CLAIRE-DSA attains excellent performance (ROC-AUC=0.98), for projection direction, overall performance is still strong (macro-AUC=0.94), with detection of oblique views appearing most challenging with ROC-AUC=0.80.
• Figure 4: Example results from the CAVE segmentation and classification. The vessel segmentation is drawn in green on the MinIP background. Note in particular the False Positive (d), where the DSA image was assessed to be suitable for the downstream task, but vessel segmentation failed, possibly due to lack of contrast medium in the image.