Additional data and information regarding:

Influence of CT Image Size and Format on Accuracy of Lung Nodule Detection

SE Seltzer, PF Judy, U Feldman, L Scarff, FL Jacobson. Radiology 1998;206:617-622.

ABSTRACT

PURPOSE:
Many radiology departments, under pressure to reduce costs, have decreased film consumption by placing larger numbers of smaller digital images on each laser film. We evaluated the impact of reducing image size on observers' ability to detect small lung nodules on chest CT scans.

MATERIALS AND METHODS:
Stimuli were 80 single slices excerpted from 13 normal chest CT cases. On one half of the images, nodules (3 to 5 mm diameter; soft-tissue attenuation corrected for volume averaging) were superimposed electronically at random locations. Four observers viewed images of six formats and sizes ranging from 6-on-1 (133 x 133 mm) to 80-on-1 (40 x 40 mm) and assigned a confidence rating about nodule presence. The images were printed using a carbon-based, dry process laser printer and were viewed under 2 protocols: (1) a fixed distance of 55 cm, and (2) an unrestricted, variable distance.

RESULTS:
At a fixed viewing distance, nodule detection decreased with smaller image sizes (more images per film). The ROC curve areas for these readings declined from 0.857 for the 6-on-1 format to 0.671 for the 80-on-1 format (p=.0001). When observers could vary their distances, the decline in ROC curve areas across all formats was not statistically significant (0.884 to 0.834; p = ns). However, there was a significant drop in performance for the smallest images (p < .05). Overall, the difference in ROC curve areas for fixed and variable viewing was significant (p < .001).

CONCLUSIONS:
Reducing image size leads to deterioration in lung nodule detection on CT scans viewed at fixed viewing distances. This can be overcome by allowing the observer to compensate for smaller images by adjusting the viewing distance.

Explanation of additional data

"Influence of CT Image Size and Format on Accuracy of Lung Nodule Detection" presented the mean performance of four observers. While a detailed presentation of each individual performance may not be of interest to the typical reader of Radiology, we have made this data available on BrighamRAD for interested investigators. This Web presentation demonstrates that the performance trends of individuals track the mean performance.

Figure 3 from "Influence of CT Image Size and Format on Accuracy of Lung Nodule Detection" presents the observers' ability to detect nodules (i.e., area under the ROC curve, A-z) on images of various format densities under the two viewing protocols: (a) fixed distance of 55 cm; or (b) variable viewing distance in which observers could get as close to the films as they felt necessary and, in addition, had access to a 3X hand-held magnifier.

Figure 3: Upper curve demonstrates results of variable viewing distance protocol with relatively constant performance with increasing image format densities. Lower curve demonstrates declining performance in the fixed viewing distance protocol. The error bars are the standard deviation of mean for all the trials.

Table of A-z for each observer

	Variable viewing distance
Format density	Obs 1	Obs 2	Obs 3	Obs 4
6	0.926	0.828	0.874	0.907
12	0.877	0.862	0.879	0.904
20	0.897	0.838	0.892	0.911
30	0.862	0.799	0.889	0.929
42	0.866	0.902	0.907	0.897
80	0.819	0.813	0.808	0.897
	Fixed vewing distance
Format density	Obs 1	Obs 2	Obs 3	Obs 4
6	0.827	0.831	0.860	0.910
12	0.749	0.853	0.887	0.923
20	0.759	0.832	0.845	0.884
30	0.659	0.797	0.891	0.844
42	0.696	0.746	0.847	0.864
80	0.613	0.644	0.685	0.743

Figure 4 from "Influence of CT Image Size and Format on Accuracy of Lung Nodule Detection" presents the fraction of nodules correctly localized to the left or right lungs on nodule images of various format densities under the two viewing protocols: (a) fixed distance of 55 cm; or (b) variable viewing distance in which observers could get as close to the films as they felt necessary and, in addition, had access to a 3X hand-held magnifier.

Figure 4: Upper curve demonstrates results of variable viewing distance protocol with relatively constant performance with increasing image format densities. Lower curve demonstrates declining performance in the fixed viewing distance protocol. The error bars are the standard deviation of mean for all the trials.

Table of fraction of correct localizations for each observer

	Variable viewing distance
Format density	Obs 1	Obs 2	Obs 3	Obs 4
6	0.838	0.838	0.863	0.863
12	0.825	0.875	0.900	0.900
20	0.850	0.838	0.913	0.900
30	0.813	0.850	0.900	0.913
42	0.863	0.825	0.900	0.913
80	0.788	0.825	0.788	0.850
	Fixed vewing distance
Format density	Obs 1	Obs 2	Obs 3	Obs 4
6	0.813	0.900	0.863	0.938
12	0.763	0.838	0.900	0.863
20	0.700	0.813	0.788	0.850
30	0.600	0.875	0.838	0.875
42	0.525	0.763	0.800	0.875
80	0.488	0.625	0.663	0.800

[Review of ROC analysis] | [Format Density]

Is this a mirrored page?
The official URL is http://brighamrad.harvard.edu/research/topics/vispercep/first.html