What's the best way to measure lung nodules at follow-up CT? Researchers from Weill Cornell Medical College say that computer-assisted markup (CAM) shows less variability than manual drawing of the nodule; accuracy is also aided by comparing more recent scans with prior scans side by side.
Much is riding on the measurement accuracy of suspicious lung nodules at CT. Substantial growth during surveillance could signal malignancy, a poor measurement technique, or something else entirely. Dr. David Yankelevitz and colleagues from Cornell sought to compare two systems for outlining nodules -- one completely manual and one assisted by the computer. The computer-assisted markup technique was developed by co-investigator Anthony Reeves, PhD, he said.
"The purpose of our study was to assess factors that affect assessment measurements other than those related to obtaining the images," Yankelevitz said in a presentation at the 2010 RSNA meeting.
The study looked at the same 12 nodules several times in different ways. The nodules ranged in volume from 29.1 mm3 to 27,700 mm3 on thin-section CT images, or from about 4 mm to 41.3 mm in diameter, he said. They were presented in random order to five readers.
Volume was measured in two ways: first, by manually outlining the boundary on each slice, and second, by using a novel method that allows the reader to choose seed points on each slice with the computer completing the outline.
The manual method calls for the placement of seed points that the reader then connects manually, he said.
"The computer-assisted method allows you to place ... a limited number of seed points in areas where there's some added difficulty; for example, a vessel or the touch of a chest wall," Yankelevitz explained. "Once that's completed, a boundary is calculated, but the reader can go back and add points or move a point if he's not satisfied and recalculate the nodule slice by slice. This allows you to leverage the best aspects of the human reader that can figure out the complex anatomy, as well as using the computer, which is better at using some predefined approaches, to define the remainder of the relatively easy boundary."
Each reading was repeated for each method, and additional readings were also performed where the readers saw their original markings in side-by-side viewing. Nodules were presented in random order each time and at least two days apart. Manual and CAM intra- and inter-reader coefficients of variation (COVs) were obtained using linear mixed effects models, incorporating random effects for reader and nodule. There were six phases in the study:
- A manual phase where each reader marked each of the 12 nodules manually
- A second phase repeating the process with the identical images in random order
- Third and fourth phases like the first and second but using the CAM technique
- A fifth phase with manual measurement but also the availability of side-by-side comparison with the prior nodule image
- A sixth phase with CAM measurement and the availability of side-by-side comparison
"When we compared the two manual reads and the two computer-assisted reads ... the computer-assisted read [showed] a decrease in inter-reader and intrareader variability," Yankelevitz said.
Inter-reader variability was greater for manual reads (COV = 11.2%) compared to the CAM technique (COV = 3.5%). Intrareader variability was also greater for manual reads (COV = 10.1%) compared to the CAM technique (COV = 6.7%).
When readers were allowed to concurrently view their earlier read side by side, the between-reader variability increased for the manual method (COV of side-by-side read = 13.9%; COVs for non-side-by-side reads = 10.4%, 12.1%).
Between-reader variability decreased for the CAM method (COV of side-by-side read = 1.4%; COVs for non-side-by-side reads = 2.9%, 3.5%).
Finally, the residual or unexplained COV decreased for the CAM method (COV of side-by-side read = 5.5%; COVs for non-side-by-side reads = 6.6%, 7.5%) and for one of the manual read sessions (COV of side-by-side read = 8.1%; COVs for non-side-by-side reads = 7.1%, 12.7%).
This result means that when each phase was modeled separately and compared side by side in the manual markup, inter-reader variability increased slightly, and for the computer-assisted technique, inter-reader variability dropped slightly, he said.
"The computer-assisted markup approach allowed improved repeatability -- that is, there was improvement in both inter-reader and intrareader variability," he said.
"These findings we interpret as suggesting that each reader when they draw something has a bias when manually determining the boundary," he said. "And when viewing side by side, the bias between the readers actually tended to increase [in the manual method] whereas it was minimized by the CAM technique."
Biases in measuring change in nodule volumes can be minimized by the use of both CAM and side-by-side viewing, he said.
