Validation of texture analysis

This thesis presents a study into the stability of texture features across multiple stages of the texture analysis pipeline. Two imaging modalities were assessed in this work: CT, in which a texture analysis phantom containing 10 texture sources was used, and PET, in which an image quality assurance phantom was used.

The stability of 43 texture features were assessed under identical imaging parameters using test-retest analysis to establish the repeatability of features in both a clinical setting, with a 10-minute rest period between scans, and texture phantom commissioning setting, with no rest period.

The effects of different CT reconstruction parameters were assessed to identify which parameters significantly impacted the image texture. A novel GLCM processing technique was developed to improve the stability of texture features across multiple reconstruction parameters. The Richardson-Lucy algorithm was applied to digital and real PET images to establish if the underlying image texture, masked by the partial volume effect, could be recovered.  

It was found that most features were stable for the rest period and no rest period scans, 95% and 93% respectively. Texture features were found to be very sensitive to reconstruction parameters, with only 14% of features being stable across the reconstruction parameters, with high levels reconstruction kernel sharpness and extreme levels of denoising being the parameters that significantly affect the most features.

With the application of GLCM processing 78% of features were found to improve in stability across the reconstruction parameters. The application of the Richardson-Lucy algorithm, with noise regularisation, was able to recover the underlying texture of the digital PET images. However, when applied to the real PET image the resultant texture was further from the ground truth than the original image.  

It is recommended that reconstruction parameters are kept identical in multi-set studies, rest periods are taken between scans, and the Richardson-Lucy algorithm is applied after feature calculation.