Session Item

Tumour segmentation is a fundamental step for radiotherapy treatment planning. To define an accurate segmentation of the primary tumour (GTVp) of OPC patients, simultaneous assessment of different image modalities is needed. Each image volume is explored slice-by-slice from different orientations, resulting in a tedious and time consuming process. Moreover, the manual fixed boundary of each segmentation neglects the spatial uncertainty known to occur in tumour delineation. This study aims to assist radiation oncologists in a slice-by-slice adaptive GTVp segmentation using probability maps, proposing a novel automatic deep learning (DL) segmentation model, on registered PET-CT images.

Based on the inclusion criteria 105 OPC patients treated with (chemo)radiation between 2014 and 2017 in our institute were included. PET and CT images and GTVp contours, used for radiotherapy treatment planning, were collected. PET and CT images were registered rigidly. Bounding boxes of 144×144×144mm³ were extracted around the oropharynx. An external validation set of 200 patients from 4 different centres was used. The DL framework was built in order to perform segmentation utilizing both inter and intra-slice context. The model was trained on sequences of 3 consecutive 2D slices of concatenated PET and CT images, while the GTVp contours are used as ground truth. A 5-fold cross validation was performed three times, training on sequences extracted from the Axial (A), Sagittal (S) and Coronal (C) plane, respectively. The Dice Score Coefficient (DSC) was used to select the best model in each fold. In the testing phase, each slice resulted in three predictions (except for slices at the boundaries of the volume) that were averaged, creating three final segmentation outputs for the A,S and C planes.

Table 1 reports quantitative results that measure the quality of the proposed model on PET-CT images. The framework has higher performance on the A and S planes compared to the C plane. The model trained on the C planes resulted in more false positives in slices without tumour. Figure 1 shows an example of a probability map obtained by the proposed slice-by-slice segmentation method. The network uses the knowledge gained by the previous and the successive slice, resulting in areas with different probabilities of predicting tumour.

Since the GTVp is used as ground truth, the quality of the contours highly affects the performance of the proposed model and the evaluation of results. The lower performance on the internal validation set, on the A and S planes, could be explained by a larger variability in GTVp contours in training set than in the external set. The bony structure and the metal artefacts on the CT images seem to be misleading for tumour classification on the C plane. The results from the proposed novel DL segmentation model are promising. The probability maps, on registered PET-CT images, can guide radiation oncologists in a slice-by-slice adaptive GTVp segmentation.