Abstract
Incidental irradiation of the heart in patients with lung cancer treated with RT produces proven cardiotoxic effects positively correlating with the absorbed dose. With the introduction of targeted therapies, the survival of lung cancer patients is increasing, resulting in higher risk to develop radiation induced cardiac toxicity. It has therefore become paramount to introduce in the clinical routine efficient methods to properly manage heart motion, which is the main reason for misdosage. In this work we introduce and validate an automatic heart segmentation algorithm on 4D CT scans, and we compare the union of its contours on all the respiratory bins with the contour of the maximum intensity projection (MIP), which is currently the most advanced clinical approach to the problem in treatment planning. 190 full hearts were manually outlined by four experienced radiation oncologists on each of ten breathing phases of the 4D CT scans of nineteen patients, who underwent lung cancer RT. The same datasets were then contoured by the automatic atlas-based algorithm using a leave-one-out approach on the 0% phases of all the patients. These contours were propagated onto the other corresponding phases for each of the patients. All the automatically generated contours were validated by the radiation oncologists, and quantitative metrics were calculated for all the cases: Dice Similarity Coefficient (DSC) and Mean Absolute Surface Distance (MASD). Moreover, for each of the patients, the union of the automatically generated contours was calculated and used as planning organ at risk volume (PRVauto). This PRVauto was then compared to the contour outlined on the MIP-CT image (PRVMIP) created with Monaco V.5.5 using the same 10 breathing phases of the 4D CT. DSC and MASD (shown in Table 1) are comparable to inter-operator variability in clinically acceptable whole heart delineations. The comparison between PRVauto and PRVMIP produced an overall mean DSC of 0.90 (range 0.84-0.93) and a mean MASD of 3.8 (range 2.2-6.2); the lower DSC was due to inconsistencies in manual contours on subsequent axial slices, resulting in jagged edges in the sagittal and coronal planes. The automatic atlas based contouring algorithm on 4D CT datasets proposed in this work produces robust individualized heart delineations that are easy to implement in the RT routine. The 4D volume of the heart created can be used as reliable PRV for thoracic RT planning. In particular for cases where the treatment target is proximal to the heart, where it is essential to guarantee the highest accuracy possible, this algorithm proved to be more reliable than the MIP based contours.Abstract 3689; Table 1Respiratory PhaseAverage DSCSDAverage MASD (mm)SD0%0.920.022.61.210%0.920.022.71.220%0.910.033.01.430%0.900.033.51.640%0.890.033.91.650%0.880.034.21.660%0.880.044.21.770%0.890.033.71.680%0.910.023.01.390%0.920.022.81.3 Open table in a new tab
Highlights
Purpose/Objective(s): To develop and demonstrate the efficacy of a novel head-and-neck multimodality image registration technique using generative adversarial networks, and show its improvement over traditional direct multimodal registration
We investigate the feasibility of using virtual non-contrast (VNC) images derived from dual-energy CT (DECT) to eliminate the pre-contrast CT and the registration error
In this work we introduce and validate an automatic heart segmentation algorithm on 4D CT scans, and we compare the union of its contours on all the respiratory bins with the contour of the maximum intensity projection (MIP), which is currently the most advanced clinical approach to the problem in treatment planning
Summary
Purpose/Objective(s): To develop and demonstrate the efficacy of a novel head-and-neck multimodality image registration technique using generative adversarial networks, and show its improvement over traditional direct multimodal registration. Materials/Methods: CT datasets, including DECT and conventional 120 kVp pre- and post-contrast CTs, acquired for 10 pancreatic cancer patients were used.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have