MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Machine Learning On Quality Control of Chest CT Chest Exams: Scan Length Optimization

D Huo*, A Scherzinger , University Colorado Denver, School of Medicine, Aurora, CO

Presentations

(Tuesday, 7/31/2018) 4:30 PM - 6:00 PM

Room: Room 202

Purpose: Tremendous efforts have been made to improve image quality and reduce exposure in CT. However, one important parameter, scan length, is less frequently optimized or even monitored. To understand how much unnecessary CT scan coverage (“over-scan�) was performed in routine clinical CT exams, we inspected all the “CT LUNG CANCER SCREENING� exams that were performed in our facility between September 2016 and September 2017. Machine learning algorithms were used to detect the lung region from the scout PA images, and “required� scan range calculated automatically. This scan range was then compared with the actual scan range.

Methods: A neural network model based on the “UNET� framework was established, and the model trained with the public dataset “JSRT�, which contains 247 chest x-ray images and corresponding lung masks. This model was then further trained with 12 CT scout images and corresponding lung masks from our local dataset. The trained model was applied to the scout images of the remaining CT lung cancer screening exams to detect and segment the lung region. The upper and lower boundaries of the segmentation maps were recorded as the “required� scan range, and then compared with the actual scan range.

Results: For the 145 exams that were included in this study, the average “required� scan length was 253 mm with standard deviation of 27 mm. The average “over-scan� was 62 mm or 25%. The upper boundary (Superior) on average has an “over-scan� of 17 mm, and the lower boundary (Inferior) on average has an “over-scan� of 45 mm.

Conclusion: Scan coverage of about 25% more than “necessary� has been observed in CT Lung Cancer Screening exams. Even allowing for some safety margin to ensure coverage, improvement opportunity exists to further reduce patient dose and radiologist workload by reducing scan length.

Keywords

Quality Assurance, CT, Computer Vision

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email