ESTRO 38 Abstract book

S612 ESTRO 38

PSMA PET/CT for monitoring of salivary gland cell loss is feasible, with acceptable patient burden, low radiation burden, and sufficient image quality for quantitative gland-based and voxel-based evaluations. Anatomical changes after treatment can complicate image registration. PO-1102 Multi-atlas vs. single-atlas auto-segmentation for Head and Neck OARs: time efficiency and accuracy D. Gugyerás 1 , A. Farkas 2 , M. Csima 3 , Z. Cselik 4 , J. Hadjiev 2 , A. Gulyban 5 , F. Lakosi 2 1 Kaposi Somogy County Teaching Hospital, Dr.József Baka Diagnostic- Radiation Oncology- Research and Teaching Center, Kaposvár, Hungary ; 2 Somogy County Kaposi Mor Hospital, Dr. József Baka Diagnostic- Oncoradiology and Research Center, Kaposvar, Hungary ; 3 Kaposvar University, Faculty of Pedagogy, Kaposvar, Hungary ; 4 Csolnoky Ferenc Hospital, Radiation Oncology, Veszprem, Hungary ; 5 Europa Hospitals, Radiation Oncology, Brussles, Belgium Purpose or Objective to test and compare time efficiency and accuracy between single-atlas (SingleA) and multi-atlas (MultiA) based auto- segmentation in Head and Neck (H&N) organs at risk (OAR) delineation. Material and Methods 60 patients with hypopharynx/larynx cancer were selected and divided into two groups: ideal population (T1-T2/N0-1, n=30 pts) and clinical scenario (≥T3/ ≥N2a, n=30 pts). All manual delineation was performed by an expert RTT, including semi-automated options ( Ref ). The baseline single- and multi-atlas for both groups consisted of 10 reference delineated cases with 20 OARs (Brain, Brainstem, Oral cavity, Mandible, Oesophagus, PharynxConstr, PituitaryGland, Spinal Cord, ThyroidGland, Trachea, L/R Submandgland, L/R Eye, L/R Lens, L/R OpticNerve and L/R Parotis). For the next 10 patients (11-20th) atlas contours were generated with both MultiA and SingleA, followed by manual corrections ( MultiCor, SingleCor ). For the remaining patients (21- 30th) the extended atlas (consisting of 1-20th cases with “Ref” contour) was used to generate the same set of OARs. Time required for Ref, Multi, MultiCor ( T_MultiCor ), Single, SingleCor ( T_SingleCor ) were measured and compared using t-test. Delineations ( Ref vs . Multi, MultiCor and Single, SingleCor ) were compared using relative Dice Similarity Coefficient (DSC), Jaccard index (JI), commonly contoured volumes (CCV) and 95% of the Hausdorff distance (HD95%). Baseline vs. extended atlas performance was compared using two sided t-test with p<0.05 significance level. Results MultiA outperformed in 57.5 % of the volumetric parameters the SingleA with significantly lower T_Cor as well (11min:19sec vs. 15:20, p<0.001) (Table1-2). By increasing the number of atlas cases a >7 min gain per patient (8:37 vs. 15:57, p<0.001) was achieved. Time saving ( T_MultiCor- T_SingleCor ) did not differ between ideal and clinical population (p=0.66) and became more pronounced with advanced atlases (4:00 vs. 7:20, p<0.001). The extended atlas did not improve delineation performance compared to baseline atlas.

minutes each. PET images were reconstructed to voxels of 2x2x2mm 3 . Low dose CT parameters included 2 mm slices and 40 mAs with dose optimization. PSMA PET images were co‐registered to the original planning CT using rigid registration, based on the accompanying low dose CT. The image quality and image registration, along with the change in uptake were assessed visually. Results At the time of this evaluation, all included patients had completed their scheduled scans successfully. 7 had received a baseline and mid-treatment scan, 5 had received a 1 month post-treatment scan and 3 had received a 6 month post-treatment scan. The radiation dose per scan was estimated at 3 mSv (1 mSv for PET, 2 mSv for CT). The image quality of the PSMA PET and low dose CT were visually good and without artefacts (figure 1). Image registration between the PSMA PET and the planning CT was visually good for the baseline scan, but was increasingly challenging during and after treatment due to anatomical changes caused by weight loss and gland regression. Areas of signal loss corresponded to the dose distribution (figure 2).

Conclusion