Computed Tomography (CT) Image Quality Analysis Methods, Noise Effect and Calibration

Abstract

Commutated tomography (CT) imaging has become more common for medical diagnosis in recent years. Since a correct diagnosis has a substantial impact on patient outcomes and treatment plans, the rapid identification of strokes using CT scans is essential for prompt clinical action.  In order to train and evaluate deep learning models, this work made use of a publically available dataset of CT pictures of brain strokes. The collection includes labelled images reflecting various stroke states.  Artefacts like as noise and intensity changes are common in CT scans and might make it difficult to classify strokes accurately. In order to overcome these obstacles, four deep learning models, CNN, VGG16, ResNet, and Multilayer Perceptron were used for noise-aware preprocessing in CT images and for stroke identification. To make the model more generalizable and resistant to imaging discrepancies, the preprocessing pipeline included noise reduction methods, intensity normalization, and data augmentation. The F1-score (F1), recall (REC), accuracy (ACC), and precision (PRE) were utilized for a comprehensive evaluation of the classification abilities. With a remarkable 99.50% accuracy, convolutional neural networks (CNNs) were able to extract spatial and hierarchical information from noisy images. ResNet came in second with 97.75%, VGG16 third with 96.50%, and MLP fourth with 96%. These results demonstrate the importance of preprocessing and noise handling in enhancing classification reliability. The proposed framework shows promise for real-time clinical deployment, supporting automated and rapid stroke detection to reduce diagnostic errors and improve patient care outcomes.