Purpose: Daily dose reconstruction in lung cancer is hampered by inaccurate tissue densities obtained from cone-beam-CT (CBCT) images, leading to dosimetric errors of ±15%. We propose to generate synthetic CT images from CBCT images in patients with lung cancer to restore the accuracy of the physical densities of thoracic tissues and thus enable accurate dose calculation for each treatment. This approach may reveal dosimetric or geometric reasons for treatment failure or excessive toxicity.
Methods: Data was collected from a retrospective cohort of patients (n=23) with non-small cell lung cancer. Planning CTs, treatment plans, dose distributions, and CBCTs at all fractions of delivered radiotherapy (2Gy/fx) were acquired, under institutional approval and informed consent. 22 datasets, consisting of the treatment planning CT and first fraction CBCT, were used for training and building the cycle-GAN model to generate synthetic CTs. A synthetic CT was generated for one validation dataset. Hounsfield units from the synthetic CT were compared with its corresponding CT. Both images were compared using histograms of Hounsfield units.
Results: A synthetic CT was successfully generated at 100 epoch. The synthetic CT had good bone and muscle definition and retain the structures within the lung. Hounsfield units are comparable to the corresponding CT images for bone, fat, muscle, and lung tissue. The histograms demonstrate an improved assignment of Hounsfield units for the synthetic CT, compared to the original CBCT.
Conclusion: We demonstrated the feasibility of cycle-GAN generated synthetic CTs from original CBCT datasets. With a training dataset of 22, the histograms of grayscale values are similar to the planning CT dataset. We are currently improving the training set to proceed with more in-depth analysis and validation of the generated synthetic CT images and doses calculated on these images.