206. SEMI-AUTOMATED STATISTICAL EVALUATION OF NUCLEI IN HISTOLOGICAL SECTIONS OF BREAST CANCER TISSUE

Department: NanoEngineering
Research Institute Affiliation: California Institute for Telecommunications and Information Technology (Calit2)
Faculty Advisor(s): Andrew Kummel

Primary Student
Name: Manuel Esteban Ruidiaz
Email: mruidiaz@ucsd.edu
Phone: 858-534-2752
Grad Year: 2010

Abstract
Background: Separating malignant from benign breast epithelial cells from a small sample of cells such as in fine needle aspiration can be a diagnostic challenge. We have developed a semi-automated system for statistical evaluation of histological sections of breast cancer. This process objectively grades suspected cancer cells and determines which quantifications of nuclear characteristics best differentiates cancer and benign cells.

Methods: Hematoxylin and eosine (H&E) stained histological sections were analyzed for nuclear features using an automated nuclear analysis system, which can outline and extract measurement parameters both from individually isolated nuclei and from groups of cells. The isolated characteristics included area, circularity, Feret's diameter, maximum, mean, minimum, standard deviation of staining intensity and perimeter. The group characteristics included the minimum local inter-nuclear distance, the local nuclear density, the fractional nuclear area, fraction of area covered by high circularity nuclei, and fraction of nuclei with high circularity. The surgical samples were classified into high grade ductal carcinoma in-situ (HGDCIS), low grade ductal carcinoma in-situ (LGDCIS), high grade invasive (HGINV), low grade invasive (LGINV), and benign. A total of 25 surgical cases and 25,000+ cells were evaluated. The images to be analyzed were reviewed by a pathologist to confirm their correct histological designation.

Results: Principle component analysis (PCA) yielded 14 distinct non-correlated variables. Enhanced separation of cancer classes with minimal overlap was achieved. The initial 5 components accounted for 29.42%, 21.01%, 14.96%, 10.96%, 8.52% of the proportion of variance. With linear discriminant analysis, 4 variables accounted for 55.92%, 31.12%, 12.04%, 0.92% of the proportion of discrimination respectively. The first variable (LDA1) had means of -11.12, -9.83, -9.26, -8.44, -5.22 and standard deviations of 1.30, 0.51, 1.38, 0.46, 0.48 for HGDCIS, LGDCIS, HGINV, LGINV, and benign respectively. Inter-group separation is readily achievable with the first variable alone. For example, for even LGINV and normal cells, the LDA1, LDA2 centroid separation was 6x the standard deviation of the distribution of LDA1, LDA2 parameters; this allows for 99% certainty in discriminating between LGINV and normal cells.

Conclusion: The enhanced ability to discriminate using automated analysis over manual techniques provides fast, accurate, quantitative, and objective data from the inclusion of individual and cell group parameters. The value of automation is the objective quantification of grading of cancer cells versus benign. In the future we will use this type of analysis to improve quantitative pathological analysis from challenging specimens in clinical settings, such as fine needle aspirations.

« Back to Posters or Search Results