[1636] Breast Carcinoma Ki-67 Labeling Index: Comparison between Image Analysis and Expert Human Scoring with Discussion about Choosing Which Areas To Analyze

Jeffrey L Fine, David J Dabbs, Kristine L Cooper, Rohit Bhargava. Univ Pittsburgh, Pittsburgh, PA

Background: Proliferation is a key prognostic factor in estrogen receptor (ER) positive breast cancer, measured by counting Ki-67 staining. This is labor intensive and can be automated by image analysis. There are no well-established criteria for exactly how to assess Ki-67 in breast tumors with image analysis. This means that current image analysis systems must be closely supervised by their human operators, preventing significant automation of this task.
Design: 74 ER positive, resected breast tumors were stained with Ki-67 per vendor procedure (clone 30-9; Ventana, Tucson, AZ). Most-active areas were identified by an expert and then scored. This expert then scored the entire tumor section. Using highly-supervised methodology, four regions of interest (ROIs) from the most active areas were scored using an FDA-cleared image analysis system (VIAS; Ventana); then four additional ROIs were added to this analysis (the additional fields were selected from active-appearing areas elsewhere in the tumors). In a subset of 11 cases, additional scores were obtained by using random ROI selection. Statistical analyses were carried out (correlation coefficient).
Results: Image analysis correlated with the expert for both entire tumor and hot spot scoring (see Table 1). There was no significant difference in image analysis scores whether hot spots or entire tumors were analyzed, but in the 11-case subset a random ROI pattern did result in significantly lower scores (score difference with hot spot 9.18, 95%; confidence interval 5.44, 12.92). In almost all cases, manual override of automatic nuclei selection within an ROI was required. Also, the system had difficulty "seeing" negative staining nuclei even when these were manually included for analysis.

Correlation Coefficients (95% confidence intervals)
 hot spot (IA)whole slide (IA)
human0.882 (0.815, 0.925)0.863 (0.788, 0.913)
hot spot (IA)n/a0.969 (0.951, 0.981)
IA: Image Analysis

Conclusions: There was good concordance between a supervised image analysis system and an expert human scorer. Although focusing on hot spots did not affect scores, random ROI selection did; therefore ROI selection can cause inter-observer variation of image analysis. Image analysis can over-estimate by selectively seeing positive nuclei, making manual supervision important. This appears to be an issue with most current systems. Future development should focus on better automation (less and/or more efficient supervision). Guidance of ROI selection should be a priority; it is only a matter of time before whole-section analysis becomes available.
Category: Informatics

Tuesday, March 20, 2012 1:00 PM

Poster Session IV # 219, Tuesday Afternoon


Close Window