Related Experiment Video
Updated: May 16, 2026
05:45
Quantitative Immunohistochemistry of the Cellular Microenvironment in Patient Glioblastoma Resections
Published on: July 31, 2017
COMPARISON OF SPARSE CODING AND KERNEL METHODS FOR HISTOPATHOLOGICAL CLASSIFICATION OF GLIOBASTOMA MULTIFORME
Ju Han1, Hang Chang, Leandro Loss
1Life Sciences Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, U.S.A.
Proceedings. IEEE International Symposium on Biomedical Imaging
|December 18, 2012
Summary
This study compared sparse coding and kernel methods for classifying histological sections. Kernel methods, including support vector machine (SVM) and kernel discriminant analysis (KDA), performed comparably or better than sparse coding for Glioblastoma multiforme (GBM) classification.
Area of Science:
- Computational pathology
- Medical image analysis
- Machine learning in histology
Background:
- Histological section classification faces challenges due to technical and biological variations.
- Sparse coding is an emerging technique for image restoration and classification.
- Classical kernel methods are established techniques for data classification.
Purpose of the Study:
- To compare the performance of sparse coding with redundant representation against classical kernel methods for classifying histological sections.
- To evaluate the effectiveness of these methods in handling heterogeneity in histological data.
- To assess the classification accuracy for different tissue types, including normal, necrotic, apoptotic, and tumor cells in Glioblastoma multiforme (GBM).
Main Methods:
- Image patches were represented using invariant features at local (Laplacian of Gaussians) and global (color space) scales.
- Dictionaries were learned using sparse coding techniques.
- Classifiers were trained using kernel methods: Support Vector Machine (SVM) and Kernel Discriminant Analysis (KDA).
- Comparative analysis was performed on histological samples of Glioblastoma multiforme (GBM).
Main Results:
- Preliminary results indicate that kernel methods (SVM, KDA) perform as well as, or better than, sparse coding with redundant representation.
- The study highlights the effectiveness of both approaches in feature representation for histological data.
- Kernel methods demonstrated robust performance despite inherent data heterogeneity.
Conclusions:
- Kernel methods are a viable and effective approach for the classification of histological sections, particularly for complex tissues like GBM.
- Sparse coding shows potential but may require further optimization to outperform established kernel methods in this context.
- The findings support the use of SVM and KDA in computational pathology for accurate diagnostic classification.