C-arm Guidance: A Self-supervised Approach To Automated Positioning During Stroke Thrombectomy

AI summary

Overview

Research area: Computer vision for medical imaging — specifically self-supervised deep learning applied to fluoroscopy (C-arm) positioning during stroke thrombectomy.

Technical level: Advanced. The paper assumes familiarity with self-supervised learning, ResNet backbones, cross-attention, transformer blocks, linear probing, and multi-class classification metrics.

Scope: The paper introduces a self-supervised framework that uses a location-regression pretext task over synthetic X-ray images to pretrain a model that classifies 20 skeletal landmarks spanning the pelvis to the head, as a step toward automated C-arm positioning in stroke thrombectomy.

What This Paper Is About

During stroke thrombectomy, an operator must manually drive a biplane C-arm from over the patient's pelvis up to the brain, exposing the patient and operator to extra radiation and consuming minutes of attention. The paper asks whether a deep learning model can understand where the C-arm currently is relative to the patient's skeleton, and which body part it is looking at, in order to eventually automate that positioning.

Key Contributions

1. A first step toward automating fluoroscopy positioning for stroke thrombectomy with broader anatomical coverage — from the pelvis to the head — a range the paper states has not previously been explored for this procedure.
2. A self-supervised framework with a positional-understanding pretext task that predicts the spatial location of an X-ray image, and that embeds patient demographic data (age, sex, cadaver height, weight) into the learned features. The authors report this pretext task significantly improves downstream classification performance.
3. A public custom GUI built to simulate C-arm operation and to annotate landmarks on CT scans and generate synthetic X-ray images via DeepDRR, usable as an annotation tool.
4. Two datasets derived from NMDID CT scans — an annotated classification set and a large densely-sampled regression set — plus code released at https://github.com/AhmadArrabi/C_arm_guidance.

Main Findings

• Regression accuracy: The model achieved a mean positional error of 4.7 mm, which the paper describes as outperforming PoseNet and reducing the error by approximately 1.2 mm. The corresponding table reports MSE (raw) of 0.0084 and MSE (mm) of 4.7989 for the proposed model versus 0.0106 and 6.0717 for PoseNet.
• Classification when retraining all layers: The model initialized from the regression pretext task reached 0.95 precision, 0.95 recall, and 0.95 F1-score, versus 0.94/0.93/0.93 with ImageNet initialization, 0.93/0.92/0.92 with random initialization, 0.91/0.91/0.91 for PoseNet, and 0.79/0.77/0.77 for ViT-base.
• Linear probing reveals the largest gap: With only the last two linear layers trained, the pretext (regression) initialization scored 0.83/0.83/0.82, compared with 0.38/0.41/0.38 for ImageNet and 0.22/0.20/0.17 for random initialization. The authors describe this as a significant gap between domain-specific and out-of-domain pretraining.
• ViT-base performed worst among the classification baselines, which the authors attribute to transformer-based models needing more data for optimal performance.
• Ablation on probing depth: Training only one linear layer instead of two led to about a 0.2 performance drop (0.62/0.63/0.60 versus 0.83/0.83/0.82), which the authors attribute to fewer learnable parameters.
• Ablation on demographics: Removing patient demographic data produced a slight decline, from 0.95/0.95/0.95 to 0.95/0.94/0.94, indicating a modest benefit from those features.

Methodology in Plain English

The researchers worked entirely with synthetic X-ray images generated from real CT scans rather than clinical X-rays. Using whole-body CT scans from the New Mexico Decedent Image Database (NMDID), which contains full-body CT scans and demographics of over 30,000 decedents, they used DeepDRR to render Digitally Reconstructed Radiographs (DRRs). They selected scans with sharper filters (labeled "BONE_") and two scan sets — one with arms crossed ("_H-N-UXT_3X3") and one with arms raised ("_TORSO_3_X_3") — to guarantee full-body coverage.

They built a Python GUI (3D visualization with Vedo, interface with Tkinter) that simulates a C-arm, letting annotators view both 3D CT views and the corresponding DRR. For the classification dataset, annotators labeled 270 decedents, 47 of them by radiology residents and the rest by a non-physician, producing 8,080 annotations across 20 landmarks: skull, two humeral heads, two scapulas, two elbows, two wrists, T1, carina, two hemidiaphragms, T12, L5, two iliac crests, pubic symphysis, and two femoral heads. The GUI recorded only translation, i.e. 2 degrees of freedom with no rotation and a fixed z-coordinate.

For the regression dataset, they used 392 cases (270 overlapping with the annotated set) and densely sampled a uniform grid over the CT volumes with points spaced 30 mm apart vertically and horizontally, yielding 145,338 images — 76,578 from arms-raised CTs and 68,760 from arms-crossed CTs.

The model first learns a pretext task: given an X-ray image and the patient's demographic statistics, predict the image's position (x, y, z) in a coordinate system whose origin is the top-right corner of the CT scan. A ResNet-34 backbone projects the image into a latent representation; the demographic vector and the image embedding are each projected into vectors of dimension D and combined through cross-attention, where the query comes from the demographics and the key and value come from the image embedding. A standard transformer block processes the result, and a skip connection from the image embedding is combined with it via element-wise multiplication to produce a refined feature. A two-layer fully connected regression head outputs the coordinates, trained with Mean Squared Error loss. The authors hypothesize this teaches the model full-body skeletal positioning.

For the downstream classification task, the regression head is swapped for a classification head, and the model is fine-tuned with cross-entropy loss, either by retraining all weights or by linear probing (freezing everything except the last two linear layers). Experiments used PyTorch, a batch size of 512 on 8 AMD Radeon MI50 GPUs for regression and 64 on a single GPU for classification, the Adam optimizer at a learning rate of 0.0001, images resized to 256 × 256, random color jittering and posterization for augmentation, and an embedding dimension D of 128. Evaluation used a test set of 54 randomly selected decedents, with no overlap between training and testing cases.

Why This Matters

Impact on research: The paper shows that a self-supervised positional pretext task built on synthetic radiographs can outperform out-of-domain ImageNet pretraining by a wide margin in the linear-probing regime (0.82 F1 versus 0.38), and that demographic conditioning can be folded into learned image features. It also extends automated X-ray positioning research beyond the isolated body parts previously targeted in orthopedic work toward a much larger anatomical range relevant to neurointervention.

Real-world applications:

• Assisting less experienced operators in low-resource settings, where the paper notes these time-sensitive procedures are increasingly performed.
• Reducing radiation exposure to both patient and operator by shortening manual fluoroscopy during C-arm travel from pelvis to head.
• Freeing the operator's attention for other critical tasks such as device preparation.
• Serving as a simulation and annotation platform, since the released GUI simulates C-arm operation on CT data.

Industry relevance: The work is directly relevant to manufacturers of fluoroscopy and biplane C-arm systems, to developers of interventional and surgical robotics and navigation platforms, and to organizations building training simulators for interventional radiology. The paper states its motivation is semi- or automated positioning that reduces time and radiation, which maps onto concrete product capabilities for angiography suite vendors.

Future Directions

• Extending the framework toward fully autonomous C-arm control, optimizing trajectories from the pelvis to the head during thrombectomy.
• Adding rotational and depth information by annotating landmarks on the source CT scans and training on a large sample of simulated X-rays taken at different C-arm angulations, since the current annotations are translation-only with a fixed z-coordinate.
• Refining and evaluating the model on clinical fluoroscopy images acquired from biplane positioning during cerebral angiography, as the present work uses only simulated X-rays.
• Evaluating different trajectory strategies to minimize radiation dose and C-arm motion during procedures.
• The authors note that clinical translation will require a 3D model robust to diverse patient populations and noisy imaging.

Target Audience

Medical computer vision and self-supervised learning researchers; interventional neuroradiologists and neurointerventional surgeons interested in procedure automation; medical device and surgical robotics engineers working on C-arm or fluoroscopy control; and graduate students looking for a worked example of a synthetic-data, pretext-task approach to a clinical positioning problem.

Authors’ abstract

Thrombectomy is one of the most effective treatments for ischemic stroke, but it is resource and personnel-intensive. We propose employing deep learning to automate critical aspects of thrombectomy, thereby enhancing efficiency and safety. In this work, we introduce a self-supervised framework that classifies various skeletal landmarks using a regression-based pretext task. Our experiments demonstrate that our model outperforms existing methods in both regression and classification tasks. Notably, our results indicate that the positional pretext task significantly enhances downstream classification performance. Future work will focus on extending this framework toward fully autonomous C-arm control, aiming to optimize trajectories from the pelvis to the head during stroke thrombectomy procedures. All code used is available at https://github.com/AhmadArrabi/C_arm_guidance