IMAGE ANALYSIS 109 20 15 10 0.5 0.7 0.9 '• .1] 1] .3 1] .5 1] .7 1] .9 •NTERPRET•VE SCORE r0.9 Figure 9. A graph of the average number of objects for each test sample from IA versus the average interpretive score for each sample. 50 ß ß • .................................................................. nv 30 .................................................................. o o 0.5 0.7 0.9 1] .'• 1] .3 1] .5 1] .7 1] .9 •NTERPRET•VE SCORE r0.9 Figure 10. A graph of the average object area for each test sample from IA versus the average interpretive score for each sample. Statistical analysis (ANOVA and Newman-Keuls) of the IA results indicates a signifi- cant difference at the 99% confidence level between test sample C (positive control) and the other three test samples (A, B, and D). No significant differences were noted

110 JOURNAL OF THE SOCIETY OF COSMETIC CHEMISTS Figure 11. A graph of the average object area for each test sample. between samples A, B, and D. These results correlate well with the results from the interpretive evaluations, where sample C was considered comedogenic. Image analysis numbers obtained from evaluations of the same slides on separate days consistently give a correlation of 0.9. Repeated measurements on the same slide, on the same day, show a variation of less than 0.10% in the data. These results indicate that IA data is reproducible over time. The only change in instrumental set-up that has an effect on IA results is the intensity of light and degree of polarization. These can be controlled by adjusting to a predetermined threshold prior to performing IA evalua- tions. When evaluating follicular biopsy slides by image analysis there are a number of factors that can influence the results. The most obvious is image processing. Follicular biopsy slides from this study were evaluated using five different IA tasks or programs, and although the numbers were different with each task (slightly larger or smaller), the data consistently gave a linear correlation greater than 0.9 when compared to the interpretive scores. This indicates that image processing is less of a factor, providing that all of the test samples are evaluated using the same task under to same conditions. Artifacts in the FB slides present another problem. One must be sure to use slides with a confluent adhesive coating. Occasionally the glue does not adhere well to the slide and part of the "sample" is lost, leaving an open area. Follicular biopsy slides cannot have any missing sections or the IA system will give artificially low numbers. Large flakes and clumps of material should be avoided, as they will cause the IA system to give artificially high numbers. In one instance, a slide from one panelist was getting very high numbers with IA, while the interpretive score was very low. On review, this may have been