Fusion of Proteomic and Histologic Image Features for Predicting Prostate Cancer Recurrence after Radical Prostatectomy.
George Lee, Scott Doyle, James Monaco, Michael Feldman, John E Tomaszewski, Stephen R Master, Anant Madabhushi. Rutgers University, Piscataway; University of Pennsylvania, Philadelphia
Background: An estimated 30% of prostate cancer (CaP) patients suffer biochemical failure (BcF) within 10 years following radical prostatectomy. While Gleason grade is currently the best predictor of CaP aggressiveness, in several instances grade alone is unable to identify candidates for BcF. Mass spectrometry (MS) methods have recently been shown to yield predictive protein biomarkers for identifying candidates at risk for BcF. However, neither MS nor Gleason grade by itself constitutes a sufficiently strong predictor for CaP recurrence. Since the histologic image and proteomic attributes represent orthogonal sources of information, combining the channels of data could yield a new fused, feature space that can be used to identify patients at risk for CaP recurrence. However, data fusion methods for combining multi-scale, multi-modal, imaging and non-imaging data are not extant. In this work we present a novel machine learning scheme for quantitatively fusing histologic image and proteomic attributes for developing a meta-classifier for identifying candidates for BcF.
Design: Quantitative descriptors of gland morphology and nuclear architecture are computed from the digitized CaP histology. Peptide measurements are performed via MS at the site of the dominant nodule on the histology. Novel image and proteomic feature extraction tools are applied to extract quantitative data from a cohort of 19 patients (10 who were relapse free and 9 who had BcF). Our data fusion scheme was then applied to combine the image and proteomic signatures.
Results: The improved separation in the fused representation in Figure 1 is supported by classification accuracy using a Random Forest classifier, showing 39% classification accuracy on the morphological image descriptors, 43% on the architectural image descriptors, 84% on the protein features, and 86% on the fused data representation.
Conclusions: We presented a fused histology, proteomic biomarker for identifying radical prostatectomy patients who are at risk for BcF. This fused classifier outperformed Gleason grade and MS derived protein expression profiles on a cohort of 20 patients. Additional validation on a larger cohort of patients is required.
Tuesday, March 1, 2011 1:00 PM
Poster Session IV # 188, Tuesday Afternoon