Technical Workflow

1 | Modeling Phase

The Modeling Phase serves as an initial stage for fusing patient information into a single model for simulation. Since patient-specific anatomy plays a significant role in the precise prediction of RFA, we incorporate fast and accurate image processing methods for segmentation and registration and allow for manual correction. Patient- and device-specific parameters, which the RFA Guardian accepts hrough its interface, complete the patient-specific model. The first step is automatic segmentation of the liver from a PrI ceCT image. The boundary delimits the region for computation and thereby increases the performance. It also serves as an important parameter for registration throughout of the RFA Guardian. Since vascular structures in proximity to a RFA probe strongly influence the heat diffusion, the RFA Guardian consecutively registers multiple ceCT PrI images into a common coordinate system using fully automatic procedures. Although optional, optimal accuracy can only be achieved by using a ceCT image each for the arterial, portal venous and hepatic phases during processing. For the remainder of the paper, we assume availability of all ceCT phases.

Usually, all ceCT scans are recorded with minimal patient movement. However, motion correction showed to be mandatory and was therefore implemented to compensate anatomical discrepancies between individual phases due to breathing.

From these registered images, the vessel trees are then automatically segmented. For difficult cases, the RFA Guardian additionally offers tools for manual correction of non-optimal results. Automatically segmenting the tumor is barely feasible. Different tumor types expose varying tissue parameters and arbitrary localization, so we resort to a semi-automatic region growing approach using user-defined seed points.

The simulation domain is then defined by creating an optimized volumetric mesh for finite element (FE) simulation, centered at the tumor. The effect of heat deposition diminishes with distance from the probe; hence, the simulation domain is restricted to a sphere of 6cm radius and exhibits decreasing resolution with increasing distance from the tumor. This enforces high accuracy computation towards the critical region around the tumor, while concurrently balancing towards improved performance in less relevant areas.

Technical Workflow

2 | Simulation Phase

The FE mesh resulting from the modeling phase forms the domain for the simulation. The workflow of this stage splits up into several branches, depending on the use-case. In a nutshell, the basic steps comprise definition of device-specific and patient-specific parameters.

Needle Definition

The user can choose between placing a virtual needle model, or segmenting and registering a real needle from PeI CT imaging. The choice depends on the specific scenario and available data, but both ultimately yield comparable input parameters.

Real Needle

From a PeI CT image, the user can segment and register a real needle. Both prospective and retrospective scenarios profit from accurately reconstructing the geometry, e.g. of umbrella-shaped probes. The consecutive registration of this geometry employs two strategies. Initially, the RFA Guardian enforces a fast rigid registration method. In many cases, this optimally matches the images from PrI and PeI scanning sessions. However, RFA needles exhibit a certain flexibility, possibly leading to deformation and deviation from the optimal shape. Moreover, previous partial liver resection can complicate the process. Therefore, rigid registration can lead to insufficient accuracy, raising the need for an additional deformable registration method to compensate for local deformation.

Virtual Needle

If the exact needle geometry is unavailable or obstructed, e.g. due to high radio opacity after Transarterial chemoembolization, the user can place a virtual needle either directly in the PrI simulation domain, or relative to a PeI image. Generally, the virtual needle geometry can be defined using the intersection point between the needle tip and the tumor and a trocar point, which is any point along the needle shaft. However, more intricate needle models, for instance from Boston Scientific and RITA, exhibit a more complex geometry. This additionally requires parameterizing the rotation around the axis defined by the trocar and intersection. If the user places a virtual needle to fit the real model in a PeI image, the same registration procedures as for real needle identification apply.

Device-Specific Parametrization

Apart from the needle positioning, device-specific heating profiles play an important role in the simulation process. The heating profiles are vendor-defined procedures, comprising duration of heating, cooldown cycles, iteratively extending umbrella-shaped needles, target temperatures, wattage, and many more. Again, the most complex procedures result from the umbrella-shaped probes, e.g. from RITA devices. The RFA Guardian implements predefined sequences, as provided by the vendors, and let the user choose the appropriate protocol. Although the RFA Guardian provides standard presets for target temperatures and power emission, the user can also modify these to correlate to the settings used during real treatment.

Patient-Specific Parametrization

Besides device-specific parameterization, measuring or estimating tissue-specific values contributes to the overall prediction accuracy. Perfusion measurements for healthy and malignant tissue are nowadays often part of the clinical routine. Other parameters, such as specific heat capacity or thermal conductivity, can often only be estimated. Nevertheless, the RFA Guardian provides interface elements for injecting these values into the simulation in case they have been measured or can be estimated accurately.

Simulating Single & Multiple Cycles

After defining the simulation parameters, the user initiates the GPU-based computation. The visualization section of the RFA Guardian continuously display the outline of the coagulated area during this process and notifies the user of completion. Then the simulation module goes into an idle state, waiting for additional input for further ablation tasks. After executing the first standard protocol, the user can perform additional heating in the same needle position with customizable duration, or conduct additional protocols after repositioning the needle. Although, usually, the number of cycles is reasonably low, unlimited, arbitrary combinations of standard and additional heating procedures are possible. For convenience, parameterization of each cycle is stored. This enables the user to replay each step of the simulated treatment and explore different strategies.

Simulating Single & Multiple Cycles

After defining the simulation parameters, the user initiates the GPU-based computation. The visualization section of the RFA Guardian continuously display the outline of the coagulated area during this process and notifies the user of completion. Then the simulation module goes into an idle state, waiting for additional input for further ablation tasks. After executing the first standard protocol, the user can perform additional heating in the same needle position with customizable duration, or conduct additional protocols after repositioning the needle. Although, usually, the number of cycles is reasonably low, unlimited, arbitrary combinations of standard and additional heating procedures are possible. For convenience, parameterization of each cycle is stored. This enables the user to replay each step of the simulated treatment and explore different strategies.

Technical Workflow

3 | Validation Phase

The real lesion shape and size only manifests in the ceCT control one month after RFA treatment. The comparison with the simulation result in PeI demonstrates accuracy of the method. In the Validation Phase, several techniques catering to the different usage scenarios aid the user in evaluating the simulation results. When comparing simulated and real treatment, the user needs to segment the true lesion from PoI images. However, the significant time-span between PrI images and PoI follow-up leads to considerably higher abdominal deformation compared to the short interval between PrI and PeI phase.

Further, the region around the coagulated area tends to shrink, an effect that appears to be strongest in non-cirrhotic livers. These factors lead to the definitive need for deformable registration procedures to match the patient anatomy. After the registration process, the user can employ several techniques for evaluating the treatment. This includes numerical metrics, including tumor coverage of real or simulated ablations, multiple simulated ablation zones from parameter sampling and multiple segmented real ablation zones during follow up imaging. While these tasks are quite different in nature, the RFA Guardian provides a unified interface for carrying out all these tasks.

In general, the IR wants to globally determine the accuracy of the simulation. In all mentioned scenarios, this boils down to computing metrics for lesion comparison. For the presented results, relative volume difference (RVD), Sensitivity and average absolute error (AAE) have been used as major indicators in terms of accuracy. In case the IR is not satisfied with the simulation result, determining the specific problematic areas usually requires tedious, time-consuming manual measurements and slice-by-slice evaluation. To remedy these issues, the RFA Guardian provides advanced visualization techniques. The implemented approach provides three consecutive stages of evaluation, including a multivariate visualization scheme for finding and investi-gating problematic regions. Moreover, the IR can use this approach for lesion tracking in PoI monitoring, or quick comparison of different simulation parameterizations.