Noise reduction of low-dose computed tomography using the multiresolution total variation minimization algorithm

Computed tomography (CT) has become a popular tool in radiologic diagnosis due to the ability of obtaining high resolution anatomical images. However, radiation doses to patients are substantial and can increase the risk of cancer incidence. Although lowering the tube current is a direct way to reduce absorbed doses, insufficient photon numbers can cause severe quantum mottle and subsequently degrade the diagnostic value of CT images. In this study, we proposed an algorithm for noise reduction of low-dose computed tomography (LDCT) based on the multi resolution total variation minimization (MRTV) method. The discrete wavelet transform was used to decompose the CT image into high- and low frequency wavelet coefficients. The total variation minimization with suitable tuning parameters was then applied to reduce the variance among the wavelet coefficients. The noise-reduced image was reconstructed by the inverse wavelet transform. The results of the Shepp-Logan phantom added with Gaussian white noise showed that the noise was eliminated effectively and the SNR in the three compartments was increased from 2.04, 20.69 and 0.09 to 19.45, 187.77 and 0.27, respectively. In the CT image of the water phantom acquired with 50-mAs tube currents, the MRTV improved the smoothness of the water compartment. The average SNR was increased from 0.14 to 0.98, which is even better than the CT image acquired by 200 mAs. In the clinical head CT image with a tube current of 9.12 mAs, the MRTV successfully removed the severe noise in the parenchyma, and SNR was increased from 0.982 to 3.452 in average. In addition, the details of the septal structure of the sinus cavity were maintained. We conclude that the MRTV approach can effectively reduce the image noise caused by the tube current insufficiency, and thereby could improve the diagnostic value of LDCT images.