Improving MRI Safety and Efficiency: Development of an AI Model for Automated Recognition and Classification of ICD Devices Using Chest X-ray Images

Purpose: Pre-MRI safety screening in ICD patients requires accurate manufacturer identification due to varying MRI compatibility. Manual chest X-ray (CXR) review is time-consuming and error-prone. We developed a publicly accessible AI web application for automated ICD detection and manufacturer classification (Medtronic, Boston Scientific, Abbott, Biotronik) from routine CXRs or screenshots, laying the foundation for direct device-specific MRI-conditional guidance from manufacturer resources.

Methods/Materials: Publicly available and online-sourced CXRs containing cardiac implantable devices were curated and supplemented with reference images of isolated ICD models for visual differentiation. A targeted dataset of 98 ICD-containing images— including 18 distinct radiographic types based on unique can shapes and internal configurations—was annotated on Roboflow with bounding boxes and manufacturer labels. A convolutional neural network object detection model was trained locally in VSCode using exported Roboflow data and deployed on Hugging Face as a user-friendly web interface for real-time CXR or screenshot analysis.

Results: On a held-out test set of 20 additional online-sourced chest X-rays (CXRs; not used in training and representing ICDs from all four manufacturers), the model achieved 82–95%+ correct manufacturer classification with high confidence scores (≥94%) for identified targets. Variation mainly occurred in lower-contrast images with less distinct device features. Pacemaker testing yielded "No ICD from supported models detected" or low confidence (33–65%), showing excellent discrimination despite visual similarities. Inference was rapid ( < 5 seconds per image) on the deployed Hugging Face platform, enabling real-time use. These results demonstrate robustness on a limited but curated dataset of key radiographic variants.

Conclusions: This proof-of-concept web tool provides rapid, accurate ICD manufacturer classification from routine CXRs, addressing a key pre-MRI safety bottleneck and potentially reducing errors/delays for radiologists and MRI technologists. Future work: dataset expansion, automated MRI-conditional safety summaries with manual links, multi-institutional validation, and PACS integration.