A model to improve low-dose CT scan images

A computed tomography (CT) scan provides detailed cross-sectional images of the human body using X-rays. With the increased use of medical CT, concerns were expressed on the total radiation dose to the patient.

In light of the potential risk of X-ray radiation to patients, low-dose CT has recently attracted great interest in the medical-imaging field. The current principle in CT dose reduction is ALARA (which stands for ‘as low as reasonably achievable’). This could be achieved by reducing the X-ray flux through decreasing the operating current and shortening the exposure time of an X-ray tube.

The higher the dose of X-rays within a specific range, the higher the image quality of the CT image. However, a greater intensity of X-rays could potentially cause more bodily harm to the patients. Conversely, using a lower dose of radiation can reduce safety risks however this would introduce more image noise, bringing more challenges to the radiologist’s later diagnosis. In this context, low-dose CT image-denoising algorithms were proposed in a number of studies towards solving this dilemma.

Although there are many models available, the task of low-dose CT image denoising has not been fully achieved. Current models face problems such as over-smoothed results and loss of detailed information. Consequently, the quality of low-dose CT images after denoising is still an important problem.

This work has sought to improve upon existing models and discover new models that could solve the low-dose denoising problem. A high-level architecture of the system is shown in Figure 1. The trained model produces denoised CT images from low-dose images, as shown in Figure 2. The models were tested at different dose levels on a custom-developed dataset obtained from Mater Dei Hospital. The best model from the tested machine learning techniques was chosen on the basis of image quality and the model’s efficiency.

Figure 1. High-level architecture of training a model for low-dose denoising
Figure 2. An example of a low-dose CT image, the output from one of the models and the corresponding full-dose image
Student: Francesca Chircop
Course: B.Sc. (Hons.) Computing Science
Supervisor: Prof. Ing. Carl Debono
Co-supervisor: Dr Francis Zarb and Dr Paul Bezzina