Cell-Graph Mining for Breast Tissue Modelling and Classification C. Çagatay Bilgin Çigdem Demir Chandandeep Nagi Bülent Yener Motivation: The most reliable way in the current practice of medicine to diagnose cancer is the pathological examination of a biopsy which has a certain level of subjectivity. To reduce this subjectivity and have a mathematical model for diagnosing cancer tissues we consider the problem of automated cancer diagnosis in the context of breast cancer tissues. Summary: This work presents a graph theoretical technique that identfies and computes quantitative metrics for tissue characterization and classification. We segmented the digital images of histopatological tissue samples having 10 X 14 magnification and 960 X 960 pixels. Then for each image we generated cell-graphs using positional coordinates of cells and surrounding matrix components. These cellgraphs have 500-2000 cells(nodes) with 1000-10000 links depending on tissue and the type of the cell-graph being used. We've calculated a set of global metrics from cell-graphs and used them as the feature set for learning. Results: We compared our technique with other learning techniques based on intensity values of images, voronoi diagrams of the cells, and the previous technique we proposed for brain tissue images. Among the compared techniques our approach gave %79.1 accuracy whereas we obtained learning ratios of %49.2, %54.1 and %75.9 with intensity based features, voronoi diagrams and our previous technique, respectively. Department of Computer Science, Rensselaer Polytechnic Institute cs-07-02
Cell-Graph Mining for Breast Tissue Modelling and Classification
C. Çagatay Bilgin
Çigdem Demir
Chandandeep Nagi
Bülent Yener
Motivation: The most reliable way in the current practice of medicine to diagnose cancer is the pathological examination of a biopsy which has a certain level of subjectivity. To reduce this subjectivity and have a mathematical model for diagnosing cancer tissues we consider the problem of automated cancer diagnosis in the context of breast cancer tissues. Summary: This work presents a graph theoretical technique that identfies and computes quantitative metrics for tissue characterization and classification. We segmented the digital images of histopatological tissue samples having 10 X 14 magnification and 960 X 960 pixels. Then for each image we generated cell-graphs using positional coordinates of cells and surrounding matrix components. These cellgraphs have 500-2000 cells(nodes) with 1000-10000 links depending on tissue and the type of the cell-graph being used. We've calculated a set of global metrics from cell-graphs and used them as the feature set for learning. Results: We compared our technique with other learning techniques based on intensity values of images, voronoi diagrams of the cells, and the previous technique we proposed for brain tissue images. Among the compared techniques our approach gave %79.1 accuracy whereas we obtained learning ratios of %49.2, %54.1 and %75.9 with intensity based features, voronoi diagrams and our previous technique, respectively.
Department of Computer Science, Rensselaer Polytechnic Institute
cs-07-02