Session Item

Head-and-neck cancer (HNC) radiotherapy is complex, requires contouring of multiple target volumes and a large range of organs-at-risk (OARs) and is followed by a time-consuming treatment planning process to obtain a near-optimal dose distributions. Because of their focus on realism through adversarial learning, generative adversarial networks (GAN) are suitable for dose distribution prediction from a treatment planning perspective. In addition, if the input data required for dose prediction can be limited, treatment planning efficiency may be further improved. Motivated by this, we explored the relationship between various combinations of input data and how well the GAN-predicted dose distribution agreed with the benchmark clinical plan.

Data from 355 HNC patients (300/20/35 train/validation/test) treated between 2013-2019 were used: (1) planning CT, (2) OAR contours (salivary glands, swallowing structures, oral cavity, spinal cord), (3) PTV structures (elective, boost), (4) body contour, (5) clinical dose distribution. CT scans were window-levelled from -300 to +200 HU, cropped and down-sampled by half to a 128x128x64 grid with a resolution of 2x2x5mm3. Generator and discriminator networks used UNet and VGG-type architectures, respectively. The GAN-loss was a linear combination of binary cross-entropy and L1 and L2 reconstruction loss. GAN dose distributions were compared to clinical dose using dose volume histograms and mean OAR and PTV dose differences with clinical plans. We investigated input combinations: (1) CT, (2) CT+PTV, (3) CT+PTV+OAR, (4) PTV+OAR, (5) PTV+body. Mann-Whitney U-tests were used to assess statistical significance.

For each respective input combination, median[IQR] parotid L1-losses were 6.9[3.9], 2.8[0.6], 2.4[1.2], 2.4[1.5], 3.1[1.7]Gy. CT+PTV+OAR resulted in the lowest median[IQR] L1-loss for the entire body (3.3[0.6]Gy) and was significantly better than CT+PTV and PTV+body (p=0.044, 0.002, respectively).No significant differences were found between CT+PTV & PTV+body, and CT+PTV+OAR & PTV+OAR. CT-only resulted in considerably and significantly (p<0.001) worse dose distributions compared to all other input combinations.

Figure 1: Clinical and GAN dose distributions. Single slices for three examples Q1-Q3 is shown, defined by the quartiles from the CT+PTV+OAR model body contour L1-loss. Abbreviations: CT: computed tomography; PTV: planning target volume; OAR: organ-at-risk;

Expectedly, by applying dose prediction with only the CT-scan, the GAN has difficulty to determine what exactly should be the PTV for each patient. Further, GAN-predicted dose distributions were found most accurate when using all available information, but only by a small margin. This suggests that the need for OAR contours for dose prediction can potentially be circumvented. In addition, CT HUs may be replaced by the body contour. Future work should focus on investigating whether produced plans are deliverable using treatment planning software.