Open Access Data and Deep Learning for Cardiac Device Identification on Standard DICOM and Smartphone-based Chest...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Medical Imaging
• Cardiology and Cardiovascular Devices

Background:

• Cardiac implantable electronic devices (CIEDs) are crucial for managing cardiovascular conditions.
• Accurate identification and classification of CIEDs on radiographs are essential for patient care and device management.
• Existing methods for analyzing CIEDs on chest radiographs may have limitations in scope or accessibility.

Purpose of the Study:

• To develop and validate a publicly accessible deep learning model for segmenting and classifying CIEDs.
• To evaluate the model's performance on both standard Digital Imaging and Communications in Medicine (DICOM) and smartphone-acquired chest radiographs.
• To enable automated analysis of various CIEDs, including pacemakers and defibrillators.

Main Methods:

• A U-Net deep learning model with a ResNet-50 backbone was trained and validated.
• The study utilized 2321 chest radiographs from 897 patients and 11,072 smartphone images.
• CIEDs were classified by manufacturer and model, with performance measured by Dice coefficient for segmentation and balanced accuracy for classification.

Main Results:

• The segmentation tool achieved a high mean Dice coefficient of 0.936.
• Manufacturer classification accuracy reached 94.36%, while model classification accuracy was 84.21%.
• The model demonstrated robust performance across both DICOM and smartphone-based chest radiographs.

Conclusions:

• The developed deep learning model accurately segments and classifies CIEDs on chest radiographs.
• The model's effectiveness on both traditional and smartphone images highlights its potential for widespread clinical application.
• This tool offers a promising advancement for automated analysis of cardiac implantable electronic devices.