ESTRO 2025 - Abstract Book

S2537

Physics - Autosegmentation

ESTRO 2025

to automatically segment GTV in HNC from PET/CT and we test its performance against manual contouring and ATLAAS [1], a Machine learning method developed using simulated and phantom based PET images.

Material/Methods: We utilized the Hector challenge dataset [2] to train a 3D DeepLabV3+ model with a ResNet-50 backbone [3]. Training was performed on an NVIDIA RTX 3090 GPU for 100 epochs. Images were pre-processed by resampling to a 1×1×1 mm isotropic resolution. We adopted the method by Myronenko et al. [4] to extract a neck region of interest (~200×200×310 voxels). Images were normalized using z-scores, with CT normalized across the dataset and PET normalized per patient. The dataset was split 80:20 for training and testing, and the training data was expanded 10 fold using augmentation techniques. We used the deep learning segmentation output as a seed for a region growing algorithm, optimizing for the median discretized PET intensity of the region of interest. Our method was evaluated on an external dataset comprising 23 PET/CT scans from the PEARL trial [5]. Results: Our method achieved similar performance to ATLAAS, with both achieving a median Dice Similarity Coefficient (DSC) of 0.72 on the PEARL test set. Figure 1 shows GTV delineation results in axial, coronal, and sagittal views. In the two example cases, our method achieved DSCs of 0.72 and 0.60, while ATLAAS obtained 0.58 and 0.92, respectively. The box plot in Figure 2 summarises the performance on the PEARL test set and highlights the lesser variability of our method across cases.

Figure 1: Visual results for two PET/CT cases from the external test set: manual (green), ATLAAS (red), DeepLabV3+- based (blue).