Development of a Real-Time Patient-Specific Computer-Generated Model of Ablation Zones

Purpose: Accurate image-guided ablation of liver tumors larger than 3 cm in diameter is compromised by the difficulty to visualize the achieved zone of ablation in relation to the target. Although existing software may project an idealized ablation zone on a computed tomography image, this project aims to develop and display an ablation zone calculated from actual patient (vessel, tissue perfusion) and procedural (power administered, tine position) parameters. The primary objective of this research was the collection of data for refinement and validation of a computer model able to predict the geometry and volume of tissues ablated during radiofrequency ablation (RFA).

Materials and Methods: A proprietary GPU-accelerated computing architecture enabled development of a computer modeling software that computes RFA lesion geometries in 45 seconds This RFA Physics Library (Accublate; NE Scientific, Boston, MA), calculates an ablation zone from the actual power applied during an ablation, accounting for variation in tine configuration, vessels larger than 3 mm within 1 cm of the target, and parenchymal perfusion. Agar phantom perfusion studies: LeVeen electrodes and an RF 3000 generator (Boston Scientific) were deployed and activated in an agar phantom with simulated vessel (water flow [10.5 cm/s] channel in the agar, 20–35 mm from the target). Ablation data were collected, and the software then simulated three-dimensional surfaces of the ablation volumes. The modeled ablation volume was sliced in same plane as agar phantom and ablation contours compared, yielding an average distance error. In vivo studies: In two 45-kg swine, 10 computed tomography (CT)–guided ablations (3-cm LeVeen electrode, RF3000 generator) were performed at various distances from vessels. A laptop CT connected to the generator collected data; power was sampled at 1-second intervals. Electrode geometry was identified by an algorithm that detected each tine and generated a finite element electrode model. Vessels larger than 3 mm diameter within 1 cm of the ablation site were manually segmented (itk-SNAP). These data were used by the software to simulate the ablations. The standard for the ablation volume was a postablation contrast-enhanced CT; six analyzable lesions were manually segmented (itk-SNAP). For each ablation site, differences between the model of the ablation zone (computer-generated vs manufacturer’s kill chart) and the actual segmented contrast CT ablation zone were compared. Analysis: Mean and maximum errors between the segmented ablation zone geometry from the postablation contrast CT imaging in vivo and the ablation geometry predicted by the software or the manufacturer’s model were generated by measuring the off-set at 10,000 points between the actual and each predicted volume to quantify the accuracy of the models. Significance was assessed by a two-sided paired sample t-test.

Results: Comparison of the Agar Phantom ablation zone and the software-generated model yielded average error of 0.9 mm (range, 0.6–1.8 mm), with a maximum error of 2.8 mm (range, 1.6–4.9 mm). The computer model accurately accounted for the effects of an adjacent vessel. Comparison of the in vivo contrast CT segmented ablation zone with the computer model yielded average errors of 1.1 mm (range, 0.9–1.5 mm) and an average maximum error of 5.2 mm (range, 3.6–8.1 mm); for the manufacturer’s chart, the average error was 2.5 mm (range, 1.9–3.1 mm), and the average maximum error was 7.8 mm (range, 6.3–10.0 mm). The differences between the average errors and average maximum errors were significant at P = 0.003 and P = 0.009, respectively.

Conclusions: Real-time computer-generated models of ablation zones that were more accurate than the manufacturer’s charts were demonstrated in vitro and in an animal model. Clinical validation trials are in progress. The addition of an accurate real-time model of an ablation zone superimposed on cross-sectional images shou