Today’s computed tomography (CT) images are affected by inaccuracies and artifacts caused using poly-energetic photon beams. Despite active research in this field, even the most advanced image reconstruction algorithms still do not provide quantitatively accurate CT numbers. We have developed a dual-energy iterative image reconstruction algorithm (DIRA) which improves the accuracy of CT numbers by modeling the material composition of the imaged object. In DIRA, image pixels of patients are typically classified into the bone and soft tissue. Bone pixels carry information about percentages of compact bone and a mixture of red and yellow bone marrow. Soft tissue pixels carry information about percentages of water, protein, and lipid.
The estimated material composition can be used for improved medical diagnosis and treatment. For instance, DIRA can be used for the determination of calcium content in the prostate gland. Such information is useful for radiation treatment planning in brachytherapy with low-energy photons; a high calcium content in the prostate changes the spatial distribution of absorbed dose since the dose strongly depends on the tissue’s atomic number. DIRA is also useful in proton radiation therapy since the position of the dose maximum is sensitive to the material composition of the patient tissues.
The advanced algorithms used in DIRA are time demanding. To shorten the reconstruction time, we develop a deep learning algorithm capable of mimicking the performance of DIRA. Such an algorithm would perform the image reconstruction and determination of the elemental composition of tissues in a fraction of time only. In this effort, DIRA is used for the generation of training data for this algorithm.
Segmentation of pelvic bones via the 3D U-Net architecture. (a) Ground truth. (b) Prediction of our algorithm. (c) 3D view of the prediction. Taken from (González Sánchez et al., 2020) under CC BY.