Session Item

Contouring the organs at risk (OAR) accurately in head and neck (HN) radiation therapy planning is crucial for reducing treatment-induced toxicity. These delineations are time and labor-consuming and often biased by inter and intra-observer variability. Automatic deep-learning (DL) based segmentation has proven to overcome the limitations of manual delineation, yielding more robust, patient-specific contours faster. Nonetheless, these algorithms have not been integrated into the radiotherapy workflow yet, mainly constrained by the need for extensive computational resources and technical experience. This study aims to compare two different DL algorithms and assess this technology's potential in HN radiotherapy.

45 CT images from HN cancer patients with manually segmented OAR (brainstem, cord, eyes and parotids) were split into training (n = 35) and test (n = 10) sets. Two different fully convolutional neural networks were trained to segment the 4 OAR. On the one hand, the nnU-Net, a method that has shown great accuracy in anatomical delineation by automatic hyperparameter configuration and 5-fold cross-validation. This complex architecture leads, however, to long training and inference times. On the other hand, a single-class DenseVNet was trained for each OAR, using as input a bounding box built from a coarse mask obtained with a multiclass 3DUnet. This network presents certain architectural advantages that result in shorter training and inference. Both algorithms were evaluated in the test set by computing the Dice Score Coefficient (DSC) and Average Surface Distance (ASD).

Figures 1 and 2 depict the evaluation of both DL algorithms on the test set. nnU-Net achieved slightly more accurate results for every OAR but required over 180 hours for training and 5 minutes for inference, both in an Nvidia RTX 8000 GPU. Instead, each instance of the DenseVNet was trained in approximately 1.5 hours on the same GPU, a total of 6 hours for all OAR, whereas the inference drops to around 70 seconds on CPU. Moreover, DenseVNet allows for retraining one class or introducing a new OAR independently, while nnU-Net must be retrained entirely.

We present the advantages and downsides of two DL algorithms for OAR segmentation in HN CT images. nnU-Net performed slightly better for every OAR, yet this superiority is only remarkable for the parotids. Training and inference times were meaningfully longer compared to DenseVNet. The aim of DL technologies should be to shrink the time spent in manual OAR delineation by providing fast but good enough segmentations still flexible to be modified by an expert without relying on too extensive computational resources. To optimize the potential of DL techniques in the radiotherapy field, the weight between accuracy and computational resources must be therefore carefully evaluated. It is likely more efficient to train and update periodically with new data a simpler network such as DenseVNet than to rely upon complex methods as nnU-Net.