A deep learning-based automatic image quality assessment method for respiratory phase on computed tomography chest...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Radiology

Background:

• Computed tomography (CT) chest scans are crucial for clinical diagnosis, with image quality assessment (IQA) being vital.
• Current IQA is manual, subjective, and prone to human errors like fatigue and bias, impacting diagnostic accuracy.
• Objective and reliable IQA methods are needed to enhance CT examination efficiency and reduce radiologist workload.

Purpose of the Study:

• To develop and validate a deep learning (DL)-based automatic IQA method for CT chest images.
• To assess the image quality of the respiratory phase on CT chest images for optimal use in patient assessment.
• To compare the performance of the DL-based IQA method against experienced radiologists.

Main Methods:

• A retrospective study analyzed 212 chest CT scans, with data augmentation used to address data limitations.
• The DL-based IQA method integrated image selection, tracheal carina segmentation, and bronchial beam detection.
• Performance was evaluated by comparing the DL-based method's scores with the mean opinion score (MOS) from four blinded radiologists.

Main Results:

• The DL-based automatic IQA method demonstrated high performance in assessing respiratory phase image quality.
• The DL-IQA method achieved 92% accuracy in assessment scores, surpassing radiologists' 88% accuracy.
• Statistical analysis showed a Kappa value of 0.75, with 85% sensitivity, 91% specificity, 92% PPV, and 93% NPV.

Conclusions:

• A DL-based automatic IQA method for CT chest images was successfully developed and validated.
• The DL-based IQA method outperformed experienced radiologists in accuracy on the independent test set.
• This automated IQA method can potentially reduce radiologist workload and minimize errors in clinical practice.