Machine Learning and Vision Transformer for CT Scanners' Calibration and Quality Assessment

In this study, we present the process and research for finding the best machine learning methodology and innovativeapproach to evaluate the image quality in Computed Tomography (CT) scanners by predicting Signal-to-Noise Ratio (SNR) andContrast-to-Noise Ratio (CNR) from low-resolution CT images of a series of phantoms. Traditional methods of Image QualityAssessment (IQA), reliant on subjective evaluation by radiologists, often suffer from variability and inefficiency. To addressthese limitations, we explored both interpretable models like the Adaptive Neuro-Fuzzy Inference System (ANFIS) and otheradvanced deep learning architectures. Initially, ANFIS combined with Gray Level Co-occurrence Matrix (GLCM) featuresyielded suboptimal results, with an R-squared value of 0.634. Experimenting with various deep learning methodologies forimproving the performance, directed us to develop a hybrid model integrating DenseNet, Vision Transformers, andreparameterization techniques, which showed that can achieve superior results with an R-squared value of 0.8892. This researchpaper focuses on searching for the optimal machine learning model and lays the groundwork for an automated tool that canoptimize imaging protocols by providing a comprehensive quality assessment of CT images in CT calibration.