ESTRO 2022 - Abstract Book

S1670

Abstract book

ESTRO 2022

PO-1884 Automated elective lymph level segmentation for head and neck cancer radiotherapy treatment planning

V. Strijbis 1 , M. Dahele 1 , O. Gurney-Champion 2 , B. Slotman 1 , W. Verbakel 1

1 Amsterdam UMC, Radiation Oncology, Amsterdam, The Netherlands; 2 Amsterdam UMC, Radiology and Nuclear Medicine, Amsterdam, The Netherlands Purpose or Objective Head-and-neck cancer (HNC) radiation treatment (RT) planning requires contouring of elective lymph nodes in the neck, however lymph levels that need inclusion depend on clinical parameters. Therefore, a general, automated segmentation approach for total elective lymph node volume cannot be used for all locally advanced HNC. Deep learning (DL) has recently provided popular means of automated segmentation. While DL for segmentation of combined elective lymph node levels has been reported, we have evaluated and compared several DL approaches for automated segmentation of individual levels. Materials and Methods Bilateral lymph levels L1-L5 were manually contoured on computed tomography (CT) scans from 60 recent HNC patients. We trained and investigated different model arrangements of a 3D patch-based U-Net and a multi-view convolutional neural network (MV-CNN), which incorporates 2-dimensional views of each orthogonal plane at multi-scale levels to classify the voxel where the planes cross. Models were used either in a one-shot manner, to segment all lymph levels directly, or in a sequential manner, where the total lymph node volume is first segmented, followed by the MV-CNN to map foreground voxels to individual levels. Spatial performances from 5-fold cross-validation models were evaluated using dice similarity coefficient (DSC) between model and manual segmentations. Results The U-Net-MV-CNN sequential model outperformed other model arrangements, with mean[median]± standard deviation DSC scores of 86[86]±3.1% for the total lymph node volume and 80[83]±15%, 84[85]±5.2%, 80[82]±7.6%, 75[79]±13%, 71[75]±14% for each level L1-L5 individually. Typical example segmentations made with this model arrangement can be seen in Figure 1. The one-shot solution by MV-CNN yielded median DSC scores of 74[77]±9.3%, 78[79]±5.5%, 70[75]±9.6%, 64[69]±15% and 68[70]±12% for individual lymph levels.

Figure 1 Example segmentations with good, average and poor typical cases are depicted in the green, orange and red boxes, respectively. Reported DSC’s result from averaging over 5 contours. Lymph levels L1-L5 are displayed in yellow, blue, green, red and cyan, respectively. The opaque voxels show the 3D-border of the manual contour. Abbreviations: DSC: dice similarity coefficient Conclusion We investigated several automated approaches for auto-contouring lymph levels L1-L5. Although substantial variation between cases persists, requiring further investigation to minimize the need for manual checking for fully automated workflows, to the best of our knowledge, this is the first study to demonstrate that accurate (median DSC>0.8) contours of individual lymph levels can be obtained using DL methods.

PO-1885 ANALYSIS OF VASCULAR AND CIRCULAR BLOOD IN RADIATION TREATMENT PLANNING: TECHNOLOGICAL OPTIONS R.M. MEIRIÑO 1 , F. Calvo Manuel 1 , J. Burguete 2 , J. Serrano Andreu 1 , J. Aristu 1 , D. Azcona 3 , M. Cambeiro 1 , M. Vidorreta 4 , J. Pascau 5 , J.M. Delgado 3 , A. Alonso 6