Abstract
Background
Breast cancer is the leading type of cancer in women. Histologically and molecularly it is a very heterogeneous disease. Despite the categorization of breast cancers into different classes, patients with the same "profile" may respond differently to chemotherapies. Thus the need for personalized treatment becomes more and more evident. Most personalized medicine is currently based on molecular characterization of extracted biopsy samples through examination of gene expression patterns. However, tumors are spatiotemporal heterogeneous and thus these gene expression patterns extracted from a specific region do not allow for a global characterization of the tumor. On the other hand, medical imaging which is a standard of care method enables the non-invasively visual representation of the entire tumor.
Radiomics is an emerging field which involves the extraction and analysis of large amounts of quantitative imaging features from medical images. These features are able to capture the phenotypic differences of tumors and represent gene signatures non-invasively. Moreover, the radiomics data could be handled as an extra source of information that can be combined with other data, e.g. pathology, blood biomarkers and genomics in order to improve personalized treatment.
Methods and Results
In previous work we presented a novel classification approach for microcalcifications (MCs) extracted from core biopsy tissue samples and digitized using micro-CT. However, no ground truth existed for the individual MCs but only for the samples. In order to overcome this issue, we proposed the introduction of a clustering step before classification. As clustering algorithms we used Kmeans and MWKmeans and as classifiers ANNs and SVMs. We concluded that our method resulted into better results in sensitivity, specificity and accuracy, compared to the state of the art.
However, the topology of an ANN, which can highly affect its functionality, cannot be easily predefined. FD-NEAT is a very promising neuroevolutionary (NE) algorithm based on the principle of that enables evolving both the topology and the weights of an ANN by means of genetic algorithms while performing simultaneous feature selection. We are now improving the previous version of FD-NEAT and having tested it on a sub-dataset of the previously described MCs dataset due to computational and time restraints. The so far results have shown improved accuracy and a better insight on the important features.
Future Work
Our plan is to build a next generation NE algorithm able to cope with high dimensional radiomics data that we will construct by extracting new image features and combining them with genomics and other biomarkers.
Breast cancer is the leading type of cancer in women. Histologically and molecularly it is a very heterogeneous disease. Despite the categorization of breast cancers into different classes, patients with the same "profile" may respond differently to chemotherapies. Thus the need for personalized treatment becomes more and more evident. Most personalized medicine is currently based on molecular characterization of extracted biopsy samples through examination of gene expression patterns. However, tumors are spatiotemporal heterogeneous and thus these gene expression patterns extracted from a specific region do not allow for a global characterization of the tumor. On the other hand, medical imaging which is a standard of care method enables the non-invasively visual representation of the entire tumor.
Radiomics is an emerging field which involves the extraction and analysis of large amounts of quantitative imaging features from medical images. These features are able to capture the phenotypic differences of tumors and represent gene signatures non-invasively. Moreover, the radiomics data could be handled as an extra source of information that can be combined with other data, e.g. pathology, blood biomarkers and genomics in order to improve personalized treatment.
Methods and Results
In previous work we presented a novel classification approach for microcalcifications (MCs) extracted from core biopsy tissue samples and digitized using micro-CT. However, no ground truth existed for the individual MCs but only for the samples. In order to overcome this issue, we proposed the introduction of a clustering step before classification. As clustering algorithms we used Kmeans and MWKmeans and as classifiers ANNs and SVMs. We concluded that our method resulted into better results in sensitivity, specificity and accuracy, compared to the state of the art.
However, the topology of an ANN, which can highly affect its functionality, cannot be easily predefined. FD-NEAT is a very promising neuroevolutionary (NE) algorithm based on the principle of that enables evolving both the topology and the weights of an ANN by means of genetic algorithms while performing simultaneous feature selection. We are now improving the previous version of FD-NEAT and having tested it on a sub-dataset of the previously described MCs dataset due to computational and time restraints. The so far results have shown improved accuracy and a better insight on the important features.
Future Work
Our plan is to build a next generation NE algorithm able to cope with high dimensional radiomics data that we will construct by extracting new image features and combining them with genomics and other biomarkers.
Original language | English |
---|---|
Title of host publication | 6ο Πανελλήνιο Συνέδριο Βιοϊατρικής Τεχνολογίας |
Publication status | Published - May 2015 |
Event | 6th Pan-Hellenic conference on Biomedical Technology - Athens, Greece Duration: 6 May 2015 → 8 May 2015 |
Conference
Conference | 6th Pan-Hellenic conference on Biomedical Technology |
---|---|
Country/Territory | Greece |
City | Athens |
Period | 6/05/15 → 8/05/15 |
Keywords
- Radiomics
- Personalized medicine
- Neuroevolution
- Cluster analysis
- Feature selection
- Classification