An investigation into the range dependence of target delineation strategies for stereotactic lung radiotherapy

The 4DCT image sets in this study were captured using a GE Lightspeed Pro 16 slice scanner (General Electric, Milwaukee, WI), with a slice width setting of 1.25 mm. At each couch position, projection data was acquired over the duration of the tube rotation plus the additional time to complete at least one respiratory cycle. Respiratory cycle data was subsequently recorded using a Real-time Position Management (RPM) system (Varian, Palo Alto, CA). Briefly, internal target motion was synchronized to external abdominal movement by capturing a reflected infrared illumination signal from a marker block to a CCD camera. The diagram shown in Fig. 2 illustrates the waveform association with GTV movement. The marker trajectory was recorded in real time by RPM software which calculates the respiratory period based on the peaks of the observed amplitudes. Reconstructed images and respiratory data were then transferred to Advantage 4D workstation (General Electric, Milwaukee, WI) where they were sorted and binned with respect to couch position and corresponding respiratory phase at approximately ten uniformly spaced intervals within the respiratory cycle, with CT_0% corresponding to the max-inhalation phase and CT_50% the max exhalation phase. From these bins, MIP and AIP image sets were also generated by selecting the maximum and average pixel densities across all respiratory phases of the 4DCT data set, respectively. Another helical image was also taken immediately and was referred to as the corresponding free-breathing (FB) scan. The process was repeated by varying the amplitude of the tumor motion to ±0.5, ±1.0, ±1.5 and ±2.0 cm, respectively.

Manual contouring of the individual GTVs was performed on CT image sets within multiple respiratory phases in order to infer the motion information using Eclipse (version 13.6) treatment planning system (TPS). To avoid any interplanner differences, all contours were segmented by one individual using the same lung window setting in all image sets. Although we are aware of the new automated feature in version 13.6 to create an ITV via accumulations of structures in 4D, the process itself would entail naming each GTV by a common identifier which circumvents discrete phase distinction when preforming DSC analysis. Hence, the benchmark ITV structure was then generated using Boolean “OR” operation to union all 10 GTV phased structures from individual CT image sets corresponding to set motion amplitude. Alternative ITVs were also created by using a single contour outlining the FB, AIP and MIP volumetric image sets. Additionally, in consideration to clinical situations for which 4DCT is not available, an augmented FB (FB-Aug) ITV was generated containing volumetric information from the FB image set fused with the GTV contours at the maximum inhale (CT_0%) and maximum exhale (CT_50%) image sets.

Planned treatment volumes (PTVs) were created in accordance to Radiation Therapy Oncology Group (RTOG) 0915 protocol [10]. In this study we will be comparing the gold standard PTV to alternative PTV’s derived from four methods which include: (1) FB, (2) FB-Aug, (3) AIP, and (4) MIP. In the first scenario, considering only the FB (helical) image set was used, we followed conventional (non-4DCT) protocol by expanding the GTV by a margin of 5 mm in the axial plane and 10 mm along the superior-inferior direction to generate the PTV. Scenarios (2)-(4) followed 4DCT protocol where the PTVs were generated by expanding a uniformly isotropic 5 mm margin from the ITV. Figure 3 illustrates the various PTV’s as derived from the 10 phase, FB, FB-Aug, AIP and MIP techniques. Measurement tools in Eclipse TPS were used to record the volume of the PTV structures.

Letting x represent the following methods of target delineation: (1) FB, (2) FB-Aug, (3) AIP, and (4) MIP, a generic PTV ratio can be expressed as:where PTV_GS is defined as the “gold standard” PTV generated via fusion of 10 phase gross tumor delineation expanded by an isotropic 5 mm margin.

In this study, we used Dice analysis because of its simplicity to preform volumetric computations using Boolean operation within the Eclipse treatment planning system. The Dice similarity coefficient (DSC) can be used to quantify the performance of image segmentation techniques [11, 12]. DSC represent the size of overlap of two segmentations divided by the total size of both objects, and is expressed as:where V_GS is the gold standard 10 phase PTV, and V_x is the compared PTV. A DSC score of 1.0 would mean the two volumes are identical, whereas 0.0 reflects no physical overlap. It is noted the PTV ratio analysis (Rx) does not take into account the physical location of the two volumes being compared for which in certain situations a large R value may have a corresponding small DSC score. To alleviate any confusion, from hence fourth we will reserve the terminology for validating contour overlap to be in “good agreement” when DSC > 0.700, as recommended by Zijdenbos et al. [13].

Seventeen patients who had previously undergone SBRT lung treatment were also included in this retrospective study. We primarily focused on lesions in the lower lungs where a range of movement up to 2 cm was recorded by analyzing tumor minima and maxima centroid-to-centroid displacements for all ten phases within the TPS by using the “move viewing planes to structure” feature. The overall tumor movement was accounted for in the medial-lateral (∆x), anterior-posterior (∆y), and superior-inferior (∆z) directions as given by the motion vector:

Again, target volumes were segmented by one individual using the same lung window setting in all image sets.

Figure 4 shows the difference of PTVs generated by various techniques for target displacements ranging from ±0.5 to ±2.0 cm. With the exception of FB, Fb-Aug, AIP and MIP methods clearly show a linear increase in target volume for increasing range of motion along the superior-inferior direction. The negative slope associated with the FB approach (DSC = 0.740 - 0.408) is attributed to the interplay between CT data acquisition and target movement. Given our phantom study was performed at a fixed oscillation frequency of 15 cycles per minute for all ranges of motion, there exists a corresponding increase in target velocity as range increases. Considering the temporal resolution of the 4DCT imaging system is constant, the increased target velocity leads to residual motion artifacts which in this case results in a decrease in measured target volume. For all ranges of motion, a summary of the corresponding PTV ratios is given in Table 1, where subsequent DSC evaluations are indicated in parenthesis. For target displacement up to 1 cm, both FB-Aug (DSC = 0.913) and MIP (DSC = 0.913) methods have a PTV ratio of 98% with 10 phase gold standard technique, representing two viable time saving techniques for ITV delineation.

Patient data was collected from 17 cases with lung lesions displacements ranging from 0.1 to 2.2 cm. In comparison to all delineation strategies, the general inclination for FB method represented in Fig. 5 (a) shows the greatest decrease in DSC score for increasing range of motion. The association of the FB 10 mm fixed superior-inferior margin generally yields an overestimation of target overlap for small ranges of motion (< 1 cm), and an underestimation for larger motions, resulting in a mean DSC = 0.719 ± 0.106. Clinical results for the AIP technique are shown in Fig. 5 (b), where DSC = 0.803 ± 0.051. Consistent with our phantom study, DSC scores for the AIP method generally decreased with increasing range of motion, and were found to typically underestimate target coverage for which the mean PTV ratio was 0.88 ± 0.13, as illustrated in Fig. 6. MIP based delineation results are shown in Fig. 5 (c) where DSC = 0.816 ± 0.060. Contrary to our phantom outcomes, MIP more often than not underestimated target coverage, with an overall ratio of the PTVs of 0.95 ± 0.15. Finally, when compared to single CT image target delineation, the FB-Aug approach as shown in Fig. 5 (d) nominally exhibited superior target overlap for all ranges of motion yielding a mean DSC = 0.881 ± 0.056. A summary of patients results are listed in Table 2.