Objective Spectral-Spatial Analysis of Random Periareolar Fine Needle Aspiration of Women at High Risk for Contralateral Breast Cancer.

Abstract

Abstract Background: Random periareolar fine needle aspiration (RPFNA) cytology coupled with the Gail risk model has been used to predict the short-term risk of breast cancer in high risk women. The cytomorphologic features of atypical epithelial cells obtained by RPFNA can be subtle and result in inter-observer diagnostic variability and decreased sensitivity. Spectral-spatial analysis (SSA) is a method for objective image analysis that uses both color and spatial information to classify features into user-defined groups. We use SSA to classify cell clusters from RPFNA specimens into objective categories and compare that result to the cytopathology interpretation, which is the current standard.Design:Cell clusters on Papanicolau stained cytology Thin Prep slides from 7 benign and 7 malignant (14 total) breast RPFNA specimen were used to generate image stacks with the CRI Nuance platform. The specimens were processed and stained in three separate cytopathology laboratories. To build the algorithmic model, image stacks were analyzed using a neural network-based artificial intelligence system now distributed commercially as the Inform system. We manually painted green and red indicating feature (malignant) versus non-feature/background (benign) cells, respectively. A diagnostic algorithmic solution was created to stratify the new images as percent pixels correctly assigned as “malignant”. The solution was tested against cell clusters from 53 high-risk RPFNA specimen stratified by an expert pathologist (CZ) into the 5 categories of benign, epithelial hyperplasia, borderline, atypical and malignant. The specimens were collected from the contra lateral breast of patients with mastectomies for invasive carcinoma. Although 14 of the 67 cases were reused, no cellular clusters used in the training set were included in the validation set. The cytopathologist diagnosis was used as the gold standard and binarized to designate malignant cases as 1 and benign 0. These were compared to the green pixel (malignant) percentage in each case processed by INform.Results: The SSA algorithm classified all 7 malignant cases concordantly with the pathologist. The remaining 60 cases were classified as benign. The ROC curve generated from the cases had an AUC of 0.974 and an accuracy of 79.1%. The sensitivity was 100% and the specificity 76.7%.Conclusions: Spectral-spatial analysis can objectively classify benign and malignant cell clusters in excellent concordance to an expert pathologist. The epithelial hyperplasia, borderline, and atypical categories were all classified as benign by this solution representing a weakness in the solution. However, since these classes are not definitive with respect to biological behavior, the algorithm was binarized as above. In the future, algorithms will be based on biologically proven classes toward the goal of more definitive classification. A mature version of this technology could allow much broader usage of RPFNA since it would no longer be solely dependent on expert cytopathology interpretation. Citation Information: Cancer Res 2009;69(24 Suppl):Abstract nr 6001.