ESTRO 2022 - Abstract Book

S1375

Abstract book

ESTRO 2022

Conclusion This work suggests that the monolithic scintillator-based detector system design has the versatility to generate proton radiographs using two unique imaging methods and with good WET accuracy. It therefore has the potential to be translated into clinics for treatment planning and patient alignment for proton radiotherapy.

PO-1593 Automatic segmentation of individual lymph nodes in head and neck cancer patients using 3D CNNs

F. Reinders 1 , M. Savanije 1 , C. Terhaard 1 , P. Doornaert 1 , C. van den Berg 1 , C. Raaijmakers 1 , M. Philippens 1

1 University Medical Centre Utrecht, Radiotherapy, Utrecht, The Netherlands

Purpose or Objective Irradiation of individual lymph nodes (i-LNs) instead of conventional lymph node levels in head and neck cancer (HNC) patients reduces the radiation dose to nearby organs at risk, potentially leading to less radiation induced toxicity. Since contouring of all i-LNs is very time-consuming, 2 convolutional neural networks (CNNs) were trained, tested and compared for the automatic segmentation of i-LNs and LN levels on MRI. Materials and Methods Multiple Dixon T2-weighted turbo spin echo (T2 mDixon TSE) MRI scans of 25 head and neck cancer patients were used for manual contouring of i-LNs and LN levels (Ib-II-III-IVa-V) as reference. The water image and the in-phase image of the T2 mDixon TSE were used as input channels. Pre-processing was done by normalization, clipping at the 99th percentile and resampling to 1 mm ³ of all images. Two 3D convolutional neural networks (nnU-net (UNet) and DeepMedic (DM)) were trained with the scans of 15 patients. During post-processing the automatically segmented LN levels were, after manual confirmation, used as a mask to select only i- LNs segmented inside the LN levels. The MRI scans of 10 other patients were used for testing both networks (Fig. 1) with manual contours as reference. Testing metrics for the LN levels included Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance in mm (HD95). For i-LNs the testing metrics were DSC, sensitivity (SEN) and positive predictive value (PPV). SEN and PPV were based on whether the predicted segmentations intersected with the ground truth segmentations. Descriptive variables were reported as median with inter-quartile range. The metrics of both networks were compared using the Wilcoxon rank test.

Results The UNet outperformed the DM network on both i-LNs and LN levels (Fig. 2). The UNet produced higher DSC scores for segmentation of i-LNs compared to DM; respectively 0.68 (0.60-0.72) versus 0.56 (0.53-0.68) (p=0.01). Comparable results were seen between both networks regarding to SEN (UNet: 0.84 (0.75-0.88), DM: 0.89 (0.84-0.96), p=0.39). The PPV was higher in favor of DM (UNet: 0.58 (0.56-0.61), DM: 0.66 (0.57-0.70), p=0.05). For most levels (II-V) on both sides of the neck the DSC and HD95 scores were significantly better with the UNet. The median DSC and HD95 score for all LN levels were 0.73 (0.70-0.76) and 6.50 (5.80-7.30) for UNet and 0.62 (0.59-0.65) and 8.29 (5.30-9.67) for DM. No difference was found between both networks in the predicted segmentations of level Ia.