Development of a No-Reference CT Image Quality Assessment Method Using RadImageNet Pre-trained Deep Learning Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate computed tomography (CT) image quality assessment is vital for diagnosis and radiation safety.
• No-reference image quality assessment (NR-IQA) is crucial when reference images are unavailable.
• Existing deep learning CT-NR-IQA methods struggle with multiple degradations and real-world variations.

Purpose of the Study:

• To develop a novel CT-NR-IQA method addressing limitations of current approaches.
• To enhance the model's ability to handle multiple degradation factors like noise and blur.
• To improve adaptability to real-world CT image degradations.

Main Methods:

• Utilized a dataset combining noise and blur to train convolutional neural network (CNN) models.
• Leveraged RadImageNet pre-trained models (ResNet50, DenseNet121, InceptionV3, InceptionResNetV2).
• Evaluated model performance using correlation coefficients between subjective and predicted scores on artificial and real clinical datasets.

Main Results:

• Demonstrated positive correlations between subjective and predicted image quality scores for both datasets.
• ResNet50 achieved the highest performance with correlation coefficients of 0.910 (artificial) and 0.831 (real clinical).
• The proposed method showed strong adaptability to real-world degradations without requiring artificially degraded images.

Conclusions:

• The novel CT-NR-IQA method effectively handles multiple degradation factors.
• Pre-trained models like ResNet50 significantly enhance performance and adaptability to real-world CT images.
• The proposed approach shows potential as a surrogate for subjective image quality assessment in CT scans.