1347743 - A deep-learning-based approach for automatic gadolinium filled catheter tip tracking in cardiac MRI guided interventions

MRI is a promising alternative to X-ray fluoroscopy for the guidance of cardiac catheterization procedures [1]. Key advantages include superior soft tissue visualization, no ionization radiation exposure to patients, and improved hemodynamic data. MR-compatible gadolinium-filled balloon-wedge catheters are used in these procedures and are tracked visually from the associated hyperintense signal of the balloon. Automatic segmentation of the catheter balloon may thus enable enhanced catheter visualization and could facilitate automated slice tracking for continuous visualization of the catheter tip during navigation [2]. In this study, we sought to evaluate the potential of deep learning for real-time automatic catheter balloon tracking.

Methods:

All imaging was performed at 1.5T (MAGNETOM Aera, Siemens Healthineers, Erlangen, Germany). A ResNet-34 network [3], pre-trained using the ImageNet dataset, was developed for binary segmentation of the catheter balloon in the images. The Dice coefficient loss function, sigmoid activation function, and data augmentation were used during training.

Training (80%) and validation (20%) data were created from real-time single shot bSSFP images acquired in 12 patients (8M and 4F, age=40±15 years) undergoing routine CMR examination (i.e. without catheter) with the following imaging parameters: TE/TR=1.25/2.5ms, Flip angle=69°, FOV=400×400mm², voxel size=1.6×1.6mm², slice thickness=10mm, BW=1002Hz/px, GRAPPA factor=2, no. slices=60 (20 in transverse, sagittal and coronal). An artificial catheter signal was retrospectively modelled as a 2D Gaussian signal and placed at various locations in the cardiovascular anatomy.  The proposed network was tested using a total of 720 images acquired from 4 patients (all male, age=10±2 years) who underwent a right heart cardiac catheterisation procedure at our institution using a gadolinium-filled balloon-wedge catheter, with a similar imaging sequence as described above and a partial saturation (pSAT) pulse (pSAT angle=30-70°) [4]. The network performance was evaluated with a binary classification confusion matrix.

Results:

Figure 1 shows an example prediction on an artificially generated input after training ResNet-34. Figure 2 highlights the predictions of the network at locating real catheter signal and tracking through different slices. Segmentation of each image took approximately 10 milliseconds. The corresponding confusion matrix is shown in Figure 3, where the sensitivity, specificity and accuracy of the network were evaluated to be 82%, 96% and 91%, respectively.

Conclusion:

A large image training dataset was artificially generated with catheter signal to successfully train a neural network to detect real catheter signal. Deep-learning-based catheter balloon segmentation is feasible, accurate and compatible with real-time conditions. In-line integration of this network is now warranted to determine the benefit of this technique for enhanced catheter navigation.