Empowering multiple instance histopathology cancer diagnosis by cell graphs
Medical Image Computing and Computer-Assisted Intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention
|December 9, 2014
Summary
This study presents a new cancer diagnostic tool combining multiple instance learning and relational learning for improved accuracy. The method leverages cell layout similarities in tissue samples to enhance automated cancer detection.
Area of Science:
- Computational pathology
- Machine learning in oncology
Background:
- Automated cancer diagnosis relies on image analysis.
- Multiple instance learning (MIL) enables diagnosis from image-level annotations.
- Relational learning can exploit spatial patterns indicative of cancer.
Purpose of the Study:
- To develop an advanced probabilistic classifier integrating MIL and relational learning.
- To improve diagnostic performance in cancer detection using tissue microarray data.
Main Methods:
- Extension of Gaussian process multiple instance learning (GP-MIL).
- Incorporation of a relational likelihood term.
- Utilizing similarity of cell layouts as relational side information.
Main Results:
- Enhanced diagnostic performance on two tissue microarray datasets.
- Demonstrated effectiveness in breast and Barrett's cancer diagnosis.
- Validation of the integrated approach for cancer pathology.
Conclusions:
- The combined MIL and relational learning approach offers superior cancer diagnostic capabilities.
- Relational information significantly improves the performance of GP-MIL classifiers.
- This method holds promise for automated cancer diagnosis in digital pathology.
