Image modality conversion from helical CT to cone-beam CT with deep learning model for adaptive radiation therapy

Radiation therapy is a common technique to treat cancer patients. The aim of each radiotherapy treatment is to deliver a sufficient radiation dose to the tumour volume in order to kill cancer cells and at the same time not hurt the healthy tissues. Radiotherapy treatments usually last from few days to several weeks. At first, to plan treatment, a computed tomography (CT) of the patient is acquired, and all the treatment is based and planned according to this initial CT, and it is assumed that the internal organs do not change position and shape throughout the whole treatment. However, during treatment, the location, shape, and size of the tumour itself can change and the location of other internal organs, making the treatment less effective. Thus, for an effective treatment, it is necessary to know the exact location at the beginning of the first exposure and before each subsequent session during the entire course of radiation therapy. Adaptive radiotherapy takes into account all the above disadvantages of conventional radiotherapy. It is a well-known technique that acquires daily images of the patient and adapts the plan to the current anatomy, ensuring that the tumour is well covered and the organs are well-spared at any time of the treatment. The use of cone-beam computed tomography (CBCT) can reduce inaccuracy in dose calculation to the tumour and surrounding organs at risk, and further constantly adjust the dose and improve image-guided radiation therapy. The advantage of CBCT images compared to conventional computed tomography (CT) is that it takes less exposure to the patient’s internal organs. Although CBCT allows the production of 3D images daily, the images suffer from severe artefacts and low contrast in soft tissue, resulting in inaccurate Hounsfield units (HU) values. These limitations make the CBCT-based adaptive planning process impractical. Therefore, it is very important to improve the image quality of CBCT. Recently, deep learning has made great strides in image-to-image conversion tasks. However, it is generally difficult to obtain paired CT and CBCT images with exact matching anatomy for supervised learning. This Master's thesis deals with the problem of producing high-quality synthesized computed tomography (sCT) images from low-quality CBCT images for prostate patients based on implementing an unsupervised deep learning method. For this purpose, a cycle-consistency generative adversarial neural network (CycleGAN) is used. The implemented CycleGAN learns the mapping between CBCT images and planning sCT images, preserving the anatomical structure from the original CBCT images and suppressing artefacts.