Session Item

Manual gross tumour volume (GTV) contouring for radiotherapy is a time and labour-intensive task. Moreover, manual contouring is inherently subjective and thus prone to inter- and intra-observer variation. Considerable inter-observer variations between manual contours are observed in about 10% of radiotherapy plans and manual target volume delineation is considered among the largest sources of uncertainty in the radiotherapy workflow. Automating GTV contouring can offer several advantages to the clinical workflow such as speed and consistency across contours. With the recent advances in semantic segmentation using deep learning, convolutional neural networks (CNN) have been shown to perform similarly to clinicians for GTV segmentations. The agreement, as measured by the Dice similarity coefficient (DSC), between automatic and manual GTV segmentations can be on par with the expected inter-observer DSCs obtained between clinicians.

A convolutional neural network can utilize information from more than one imaging modality by including several input channels to the network. In the case of PET/CT, the network can therefore incorporate the complementary and synergistic information of the metabolic and morphological tissue properties inherent in the PET and CT modalities. Recent segmentation challenges based on PET/CT, such as the HEad and neCK TumOR (HECKTOR) 2021 challenge, show that most deep learning segmentation models can achieve similar and high segmentation performance despite using different network architectures and optimalization approaches. Input image pre-processing, such as windowing of the CT channel, definition of the region of interest, and image augmentation for increased image variations can, however, have significant impacts on network performance.

This presentation will show comparisons between automatic GTV segmentations obtained using different CNN architectures, optimization approaches, and image pre-processing techniques. Single modality (CT or PET only) models will be compared to multimodality (PET/CT) models to emphasize the importance of including the information available in both modalities to obtain highly accurate and refined auto-segmentations. Automatic GTV segmentation of head and neck cancer using PET/CT will be used as an example due to the complex anatomy of the head and neck region and the frequent involvement of lymph nodes. Issues such as training, validating and testing on a single data split rather than using a cross-validation approach and the importance of an external test set will be discussed. Common errors made by the network, such as failure to delineated GTVs with low FDG uptake (false negatives) and including regions with diffuse and high FDG uptake (false positives) will be illustrated. Lastly, performance metrics and methods for incorporating patient clinical information into the models will be discussed.