2 - Computer vision segmentation of nitinol guidewires during MRI-guided cardiovascular catheterizations

Background: MRI guidance of cardiovascular catheterization procedures provides soft tissue visualization while avoiding ionizing radiation. However, metallic interventional devices such as guidewires are prone to RF-induced heating. Recently, we modified a commercial MRI scanner to operate at 0.55T to eliminate RF-induced heating, while maintaining imaging performance for real-time image guidance [1]. At 0.55T, RF power is reduced by 7.5-fold compared to 1.5T, and the burden of safety is on insufficient visualization of devices. Here, we use an interleaved bSSFP sequence producing an anatomical image and a positive contrast guidewire image [2, 3], and demonstrate a computer vision algorithm for segmentation of the guidewire.

Methods: Experiments were performed on a prototype 0.55T MRI system (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany). Animal experiments were performed according to local ethical guidelines. A 0.035"/260cm nitinol guidewire (Glidewire, Terumo, Tokyo, Japan) was used for left heart catheterization in swine. A sequence interleaving standard bSSFP imaging and positive contrast “guidewire” imaging, using low-flip-angle bSSFP and through-plane dephasing, was implemented (TE/TR = 2/4ms, flip angle = 45°/25°, matrix = 128x128, FOV = 300mm x 300mm, slice thickness=8 mm, dephasing factor = 50). Pairs of anatomical and guidewire images were acquired with 1s temporal resolution.
Previously developed threshold-based techniques [2] were insufficient to isolate guidewire signal due to the reduced susceptibility of paramagnetic guidewires at 0.55T. A mixture-of-Gaussians based clustering and region-growing segmentation was implemented using MATLAB2019a to identify the guidewire (Figure 1). The algorithm used four image inputs: the anatomical image, the positive contrast image, a subtraction image from baseline, and a frame-by-frame temporal subtraction image.  One mask was generated from the two subtraction images to identify guidewire motion. A second mask from the positive contrast images was used to isolate signal from guidewire shaft. The masks were used as seed points for region-growing, and a connected-component analysis was used to select the guidewire pixels.

Results: Figure 2 shows the anatomical image, positive contrast image and guidewire signal overlay after segmentation. The guidewire image provided consistent shaft visualization. The tapered metallic tip of the guidewire is inconspicuous on both the anatomic and guidewire images, therefore was more reliably detected using temporal subtraction images.  The combination of multiple image inputs improved the performance of the previous model.

Conclusion: At 0.55T, some commercial nitinol guidewires do not generate RF-induced heating, which may enable complex procedures under MRI guidance using standard tools. The algorithm presented here improves the in-vivo tracking of the device using multiple image types as inputs.