Development of machine learning-based optimization methods for reducing health risks in x-ray imaging and therapy applications

Abstract

X-ray is widely used for acquiring medical diagnostic images and treating cancers in non-invasive way. While using the X-ray for medical purposes, it is important to minimize dose delivered to the normal cells in the patients to minimize radiation damage. There have been researches to develop such methods. In this study, we utilized deep learning methods, which have shown outperformance in various fields including medical imaging and diagnosis, to develop methods minimize exposure to patients while obtaining computed tomography (CT) or treating patients with radiotherapy. Sparse sampling scheme is one of methods to minimize radiation dose to patients. The method reduces radiation dose by reducing number of measurements. Reducing number of measurements make the image reconstruction problem as ill-posed and introduces streak artifacts in the reconstructed images. In this study, we developed deep-neural-network based method to minimize the radiation dose with sparse sampling scheme and maintain the quality of reconstructed images. We trained neural networks to synthesize unmeasured projection data from the measured projection data. Using the network, we were able to reconstruct images from a quarter of full sampled projection data with similar quality to the image reconstructed from the full sampled projection data. The proposed method showed outperformance to the analytic interpolation methods and an iterative reconstruction algorithm. The proposed method was able to obtain high quality images with smaller computation cost and time compare to the iterative reconstruction algorithm. The radiotherapy delivers high dose to cancer cells to treat them. The treatments should be planned and evaluated before delivered to patients to optimize dose delivery. In the treatment planning process, it is important to minimize dose delivered to the normal cells, since high dose is critical to normal cells. The quality of the treatment plans for a single patient may vary with planners’ skills and experiences, since the treatment planning process high depends on planners. And the planning process is time consuming, since one need to find optimal dose delivery in an iterative fashion. In this study, we developed deep-neural-network based method to generate fluence maps, which is essential component in radiotherapy. The developed method was able to generate fluence maps without human interactions and less than a second. The generated fluence maps had similar quality to clinical fluence maps generated by expert planners. The method would helpful to reduce time required for treatment planning process and assist planners or automate treatment planning process.