ESTRO 2024 - Abstract Book

S3757

Physics - Image acquisition and processing

ESTRO 2024

Notably, the anatomical variation between the pediatric and adult populations does not seem to impact the generation of sKVCT. The metrics exhibit consistent trends across all age groups, as depicted in the figure 2 for MAE. The same trends are found for the other three metrics, PSNR, RMSE, and SSIM. When the age group of the training data is further limited to the two center quartiles, specifically the age range of 54 to 69 years, the results based on the four metrics remain consistent across all age groups. This indicates the model's capability to generalize its performance to age groups that were not part of its explicit training.

Figure 2: MAE metric per Age Groups and Sex for the adult population model

Conclusion:

The various models trained on distinct sex and age groups demonstrate promising results based on the applied metrics. These findings indicate that the models exhibit a noteworthy ability to generalize the generation of H&N skVCT scans to demographic groups that were not originally encompassed in their training data. This generalization encompasses both sexes and age factors, highlighting the robustness and versatility of the models in providing accurate and reliable skVCT images for a wide spectrum of patients, even those not specifically covered during training. The next step is to generalize the models to the five other anatomical regions in the database containing 4,000 patients.

Keywords: AI, model generalizability, synthetic CT

References:

[1] Branimir Rusanov et al. “Deep learning methods for enhancing cone-beam CT image quality toward adaptive radiation therapy: A systematic review”. In: Medical Physics 49.9 (2022), pp. 6019–6054. issn: 2473-4209. doi: