Interventional CT

Interventional radiology comprises minimally-invasive procedures such as stenting and aneurysm coiling which are based on the insertion of interventional materials like guide wires, stents or coils directly into the vasculature of a patient. Today's interventions are usually guided using conven-tional projective fluoroscopy (2D + Time) in single- or bi-plane systems. Due to the projective na-ture of this guidance the images show only a superposition of the patient's anatomy and therefore the visualization of more complex structures and their spatial relationship is often ambiguous. Such projective data may be supported with 3D information by occasionally acquiring tomographic imag-es using modern C-arm systems (Figure 1). Continuous acquisition and conventional reconstruction of CT data sets enables 4D (3D + Time) intervention guidance but, unfortunately, such fluoroscopic tomography exceeds acceptable X-ray dose levels.
In intervention guidance, multiple repetitive scans of the same body region are performed. This enables the option to incorporate prior information into reconstructions for 4D intervention guid-ance. The dose for 4D intervention guidance can be considerably reduced by combining a high qual-ity prior image without interventional materials with the special features of 4D intervention guid-ance:

• imaging high contrast structures is of the foremost interest,
• differences of consecutive data sets (time frames) are sparse,
• certain artifacts in the volume remain tolerable.

Our research group therefore developed several reconstruction methods that have the potential to be applied in real-time at very low dose values. We first used classical tools to develop the PrIDICT algorithm, that drastically reduced x-ray dose but still requires the 8-fold dose of conven-tional fluoroscopy, excessive compute resource, and a dedicated scanner with up to 16 x-ray sources. With the advent of deep learning we lowered the number of projections needed to just four at only two-fold dose-levels requiring a scanner with "only" four x-ray sources. In the future we are planning to make further use of deep learning, e.g. for motion prediction, to allow scans at the same dose levels as conventional fluoroscopy with a system that needs only a single x-ray source and detector.
The following sections introduce the concept of tomographic fluoroscopy as well our past and fu-ture work.

To help understand our aims of achieving a 3D real-time fluoroscopy at dose values as low as in 2D fluoroscopy and how corresponding medical devices may look like, we prepared a little video that illustrates our idea of how interventions may look like in the future (Figure 2).
Video

Initially, our Division of X-Ray Imaging and CT developed the reconstruction method Prior Image Dynamic Interventional CT (PrIDICT), complying with the special features of interventional guidance and combining undersampled data sets during the intervention with high quality prior data sets acquired before intervention.
This algorithm takes into account that difference images are ad hoc sparse in image domain and do not necessarily require additional sparsifying transforms. Using this reconstruction technique, un-dersampled projection data (e.g. 16 projections per half rotation) acquired at a very low dose, are sufficient to reconstruct high quality time frames based on the prior images and therefore realize 4D intervention guidance. Figure 3 shows an in in-vivo animal experiment imaged with an experi-mental cone-beam CT setup with dose levels only about eight-fold of what projective fluoroscopy requires.
Limitations: The dose levels are low compared with regular CBCT scans, but still about eight times as high as a conventional fluoroscopic intervention would require. Further on, the computation times are enormous, in particular since the running prior technique, that adapts for patient motion, is computationally demanding. Moreover, the method assumes interventional devices to move much slower than the time it takes to acquire the 16 projections. In other words: to make the method work with typical instrument motion a CT system that simultaneously acquires 16 projections would be required, i.e. a system with 16 x-ray sources and 16 flat detectors.
Video

To overcome the dose accumulation issue of PrIDICT, we recently proposed a deep learning-based method, referred to as the deep tool extraction (DTE) and the deep tool reconstruction (DTR). For each acquired x-ray projection, DTE extracts interventional tools from the patient background, and thus eliminates the need for a patient prior and a registration step altogether. The extracted tool projections are then reconstructed from 4 views only and thus a very high temporal resolution is achieved. Finally, the DTR, a CNN that learned a mapping from the corresponding 4-view CT recon-struction to a clean volumetric representation of the interventional tools, is applied.
The complete pipeline is depicted in Figure 4a and consists of firstly, extracting interventional tools in the projection domain using DTE, secondly reconstructing using the Feldkamp-David-Kress algo-rithm (FDK), and thirdly segmenting the interventional tools using DTR. Figure 4b shows reconstruc-tion results of one guide wire, one stent, and one coil reconstructed from four projections using our method.
Limitations: The dose levels are a factor two to three higher than those in conventional fluoroscopy. Similar to PrIDICT, the 4-view DTE/DTR method assumes the interventional device to move much slower than the time needed to acquire the four projections, requiring the system to be comprised of 4 x-ray tubes and detectors.