Related Experiment Videos
Training a cell-level classifier for detecting basal-cell carcinoma by combining human visual attention maps with
Germán Corredor1, Jon Whitney2, Viviana Arias3
1Universidad Nacional de Colombia, Computer Imaging and Medical Applications Lab, Department of Medical Imaging, Bogota, Colombia; Case Western Reserve University, Center of Computational Imaging and Personalized Diagnostics, Department of Biomedical Engineering, Cleveland, Ohio, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|April 7, 2017
Summary
A new computational model integrates low-, mid-, and high-level image data, including pathologist navigation patterns, to predict basal-cell carcinoma likelihood in whole slide images. This approach shows improved accuracy over traditional methods relying solely on visual features.
Area of Science:
- Digital pathology
- Computational histomorphometry
- Machine learning in oncology
Background:
- Current computational histomorphometry often overlooks expert knowledge.
- Integrating high-level information from pathologist interactions is crucial for improving diagnostic accuracy.
Purpose of the Study:
- To develop and validate a computational model (M_im) that combines low-, mid-, and high-level image features for predicting cancer likelihood in whole slide images.
- To compare the performance of M_im against a baseline model (M_ex) that relies solely on expert-labeled visual features.
Main Methods:
- M_im was developed using handcrafted low- and mid-level image features and implicitly captured high-level information from pathologist navigation data.
- A support vector machine classifier was trained on image patches to predict cancer presence.
- The model was validated on unseen fields of view from basal-cell carcinoma cases, with distinct training and testing sets to ensure patient independence.
Main Results:
- M_im achieved an accuracy of 74.49% and an F-measure of 80.31%.
- The baseline model (M_ex) achieved an accuracy of 73.47% and an F-measure of 77.97%.
- M_im demonstrated superior performance compared to M_ex in predicting cancer presence.
Conclusions:
- Computational models integrating diverse image information, including expert navigation, can enhance cancer detection accuracy in digital pathology.
- The developed M_im model shows promise for improving the efficiency and reliability of basal-cell carcinoma diagnosis from whole slide images.