Recent advances in minimally invasive vascular disease treatments have led to the use of interventional tools like guidewires and stents,1 guided by fluoroscopy with high temporal resolution but limited depth information. To address this limitation, there is a growing interest in 3D image guidance, or 4D interventional guidance, which involves displaying a series of 3D images during procedures. However, implementing X-ray-based 4D interventional guidance requires a high-temporal-resolution reconstruction algorithm with minimal dose per 3D reconstruction. V¨oth et al.2 proposed, based on prior work of Eulig et al.,3, 4 an algorithm for the 3D reconstruction of interventional material from only two newly acquired X-ray images. Their pipeline utilizes the deep tool extraction (DTE) algorithm to compute interventional material images, which are then back-projected into a volume. A 3D U-Net5 called the deep tool reconstruction (DTR) transforms these backprojections into 3D reconstructions of the interventional material. While the pipeline shows impressive 3D reconstruction quality, it occasionally outputs false positives or negatives. In this work, we enhance the temporal information utilization by feeding the reconstructions of previous time steps as additional inputs to the DTR, improving the Dice coefficient from 71.21% to 76.84% on a simulated guidewire dataset.
|