Lymph Node Metastasis Status in Primary Breast Carcinoma Can Be Predicted Via Image Analysis of Tumor Histology
Mark D Zarella, Md Alimoor Reza, Yulan Gong, David E Breen, Fernando U Garcia. Drexel University College of Medicine, Philadelphia, PA; Drexel University, Philadelphia, PA
Background: Axillary lymph node metastasis status remains one of the most critical prognostic variables for breast cancer management and patient survival. Methods to improve reliability must be developed to avoid unnecessary surgeries and complications. The objective of our study is to demonstrate that lymph node metastasis status may be predicted via computerized image analysis of primary breast tumor histology.
Design: High-resolution (0.5μm/pixel) whole-slide images were produced from primary breast carcinoma specimens stained with a complete prognostic panel. Cell structures were extracted from these images using a custom segmentation paradigm. The properties of these structures were analyzed using stochastic geometric and chromatic transformations, forming a set of feature distributions that characterized the attributes of the cells in each sample. We constructed predictions individually for each feature distribution using a support vector machine classifier (SVM), and formed a final prediction from the individual predictions in an adaptive manner. Each prediction was accompanied by a confidence score, allowing us to relate the accuracy of the prediction to the certainty of the classification.
Results: We found that this procedure is highly predictive of axillary lymph node metastasis. The system achieved >95% correct rate for the highest scoring cases, which comprised over half the total number of cases we evaluated. Analysis of these cases revealed that N0 predictions correlated with T1 tumors (<2cm) of histologic Nottingham grade (HNG) 1 with low Ki-67 scores (<10). N1 predictions revealed T2/3 tumors of HNG 2/3 and high Ki-67 scores (>20).
Conclusions: We have demonstrated a fully automated procedure to predict metastasis status from histopathological images.
Monday, March 4, 2013 11:00 AM
Proffered Papers: Section H2, Monday Morning