Learning the topological properties of brain tumors Çigdem Demir S. Humayun Gultekin Bülent Yener Image representation machine learning medical information systems Different types of feature representation have been investigated to represent the histopathological images for the purpose of cancer diagnosis. In this work, we demonstrate that cell-graphs provide effective representations as they encode the pairwise relation between every cell by statistically assigning a link between them. Working with photomicrographs of 646 archival brain biopsy samples from 60 patients, we show that without this pairwise relation, neither the spatial distribution of the cells nor the texture analysis of the images yields as accurate results as in the case of the cell graphs to distinguish cancerous tissues from non-cancerous tissues with similar cellular density levels. We use the global graph metrics that are defined on the entire cell-graph as a feature set of a multilayer perceptron for the tissue level diagnosis of a brain cancer called malignant glioma. In our experiments, we correctly classify the cancerous and healthy brain tissue samples that have significantly different cellular density levels with accuracy greater than 99 %. Furthermore, we accomplish distinguishing the cancerous tissues from non-neoplastic reactive/inflammatory conditions that may reveal an equally high cellular density; with an accuracy of at least 92 %. Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY cs-04-14
Learning the topological properties of brain tumors
Çigdem Demir
S. Humayun Gultekin
Bülent Yener
Image representation
machine learning
medical information systems
Different types of feature representation have been investigated to represent the histopathological images for the purpose of cancer diagnosis. In this work, we demonstrate that cell-graphs provide effective representations as they encode the pairwise relation between every cell by statistically assigning a link between them. Working with photomicrographs of 646 archival brain biopsy samples from 60 patients, we show that without this pairwise relation, neither the spatial distribution of the cells nor the texture analysis of the images yields as accurate results as in the case of the cell graphs to distinguish cancerous tissues from non-cancerous tissues with similar cellular density levels. We use the global graph metrics that are defined on the entire cell-graph as a feature set of a multilayer perceptron for the tissue level diagnosis of a brain cancer called malignant glioma. In our experiments, we correctly classify the cancerous and healthy brain tissue samples that have significantly different cellular density levels with accuracy greater than 99 %. Furthermore, we accomplish distinguishing the cancerous tissues from non-neoplastic reactive/inflammatory conditions that may reveal an equally high cellular density; with an accuracy of at least 92 %.
Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY
cs-04-14