CT is a versatile medical imaging method to diagnose and monitor patient diseases. However, varying patient characteristics and scan settings create challenges in maintaining consistent image quality, complicating image comparisons, especially across different sources. The reconstruction kernel in CT image reconstruction is a key parameter in the reconstruction process. It affects image characteristics, such as sharpness, contrast, and noise. There is an urgent need for a method that effectively compares and categorises reconstruction kernels from different vendors using real patient scans. Therefore, this thesis focuses on extracting features from real patient images to facilitate kernel comparisons within and across manufacturers.

This research aims to create a machine learning (ML) method that categorises reconstruction kernels from various vendors into groups based on their sharpness. This categorisation will rely on image features extracted directly from real patient scans with diverse scan parameters.

Two distinct methods were explored to achieve the objective, each utilising different image features and applied to a selected subset of the CT datasets from the National Lung Screening Trial (NLST) and the Lung Image Database Consortium image collection (LIDC-IDRC). The first method focused on noise features, specifically the standard deviation (SD) of the most homogeneous region of interest (ROI) to measure CT scan noise magnitude and the central frequency (CF) derived from the noise power spectrum (NPS) to represent scan noise texture. These noise features were used as input for a linear support vector machine (SVC), creating the $SVC\_noise$ model. Additionally, an approach that incorporated radiomic features was explored. These radiomic features were extracted from 30-pixel-sized ROIs selected from the ten most homogeneous patches. The radiomic feature sets were then used to train a random forest classifier (RFC), creating the $RFC\_radiomics$ model. The models were evaluated using accuracy and Receiver Operating Characteristic Area Under the Curve (ROC AUC) scores. McNemar’s test was employed to determine if one model significantly outperformed the other. Evaluating the categorisation results presented a significant challenge due to the lack of a ground truth. Consequently, a subset of the smoothest and sharpest kernels from each manufacturer was selected to train, validate, and test the models. Subsequently, the models were applied to the remaining kernels, and ground truth was established for each kernel by identifying the predominant class within each one.

Both models demonstrated strong performance when applied to 270 cases featuring 37 distinct reconstruction kernels. The $SVC\_noise$ model achieved an impressive ROC AUC score of 0.97 and misclassified eight of the 270 cases based on its smooth and sharp categorisation definition. The $RFC\_radiomics$ model achieved a slightly lower ROC AUC score of 0.96, with ten misclassifications out of the 270 cases. McNemar’s test indicated that the difference in performance between the two models was not statistically significant. Moreover, the ground truth approach, applied manually, resulted in only one inconsistent kernel between the two models; specifically, the determination of the ground truth of kernel “$B50s$” differed.

In summary, the $SVC\_noise$ and $RFC\_radiomics$ models displayed promising performances, with neither significantly surpassing the other. Both models exhibited the capacity to effectively identify sharpness-related patterns within the two classes while disregarding the noise caused by variations in scan parameters and patient characteristics in real patient data. This capability offers valuable insights that can bridge the divide between research and clinical applications. However, it is important to note that the findings from this research are preliminary, and caution should be exercised when applying these results to broader contexts, including newer reconstruction kernels and techniques.