Session Item

Proton dose calculations based on 4DCBCTs require image intensity corrections. This retrospective patient study compares the dose accuracy of a scatter-corrected 4D cone beam CT (4DCBCTcor), generated in a computationally expensive workflow, which is based on a 4D virtual CT (4DvCT) prior obtained from deformable image registration, to a deep learning-based method adapted to 4DCBCT, called ScatterNet (SN).

For 26 lung cancer patients, treated with photon therapy at the LMU University Hospital, planning CTs (pCTs) and corresponding CBCT projections were available.  The ScatterNet U-shaped convolutional neural network (CNN) architecture was trained with paired raw and corrected CBCT projections, with the latter ones generated within the 4DCBCTcor workflow.  The network was trained in 2D to perform corrections in the projection space. The number of patients with a total of 17,564 2D images was split 15, 6, and 5 for training, validation, and testing, respectively. The reconstruction method MA-ROOSTER with the same settings and vector fields was used for both the corrected projections from the network and from the conventional workflow, yielding the 4DCBCTSN and 4DCBCTcor.

Using the treatment planning system RayStation, robust intensity-modulated proton therapy (IMPT) plans administering 8 fractions of 7.5 Gy with a 3-field arrangement were created on free-breathing pCTs, contoured by a trained physician. A density override of the internal target volume (ITV) with muscle tissue was performed. On every breathing phase of 4DvCT, 4DCBCTcor, and 4DCBCTSN the proton dose was recalculated without density override. For the test patients the dose was quantitatively analysed using dose-volume-histograms (DVH) parameters and gamma pass-rate analysis, using the 4DvCT as ground truth. The results of the gamma analysis will be presented at the conference.

Both 4DCBCT correction methods showed substantial and similar image quality improvements over the initial uncorrected 4DCBCT. Table 1 shows the high dose agreement for the different modalities in terms of ITV D98% among the test patients. The median differences were no larger than 1.2%, 1.9%, and 1.1% between CBCTcor-vCT, CBCTSN-vCT, and CBCTSN-CBCTcor, respectively. Figure 1 presents dose difference plots between CBCTcor-vCT, CBCTSN-vCT, and CBCTSN-CBCTcor with small deviations of a few percent around the ITV, however larger deviations were observed in lung tissue.

Phase-dependent proton dose calculation for lung cancer patients is feasible using ScatterNet for 4DCBCT intensity correction. IMPT plans based on network images showed comparable quality to those of the non-deep learning 4D scatter correction workflow. A substantial reduction in computational time (scatter correction for all phases per patient decreased from more than 10 min to a few seconds) makes this CNN method potentially interesting for daily clinical decisions such as the necessity of replanning.