Novel Mahalanobis-based feature selection improves one-class classification of early hepatocellular carcinoma
Ricardo de Lima Thomaz1, Pedro Cunha Carneiro2, João Eliton Bonin3
1Biomedical Engineering Lab, Faculty of Electrical Engineering, Federal University of Uberlândia, Av. João Naves de Ávila 2121, Uberlândia, MG, 38408-100, Brazil. rlthomaz@outlook.com.
Medical & Biological Engineering & Computing
|October 17, 2017
Summary
Early hepatocellular carcinoma (HCC) detection improves survival. A new algorithm using multi-objective Mahalanobis fitness effectively selects features for one-class classification of early HCC from CT scans, achieving 0.84 AUC.
Area of Science:
- Medical Imaging
- Machine Learning
- Oncology
Background:
- Early detection of hepatocellular carcinoma (HCC) significantly increases survival rates.
- Multidetector computed tomography (MDCT) is crucial for HCC detection.
- Identifying optimal features for early HCC classification in MDCT remains a challenge.
Purpose of the Study:
- To introduce an unconstrained genetic algorithm (GA) feature selection method for early HCC detection.
- To enhance the performance of one-class classifiers in identifying early HCC.
- To determine the most relevant features for early HCC characterization in MDCT.
Main Methods:
- Developed an unconstrained GA feature selection algorithm with a multi-objective Mahalanobis fitness function.
- Compared the proposed algorithm against constrained Mahalanobis functions and other unconstrained methods.
- Evaluated classifier performance using cross-validated one-class Support Vector Machines (SVMs).
Main Results:
- The proposed multi-objective Mahalanobis fitness function significantly reduced data dimensionality by 96.4%.
- Achieved improved one-class classification performance for early HCC with an Area Under the Curve (AUC) of 0.84.
- Identified intensity features from arterial to portal and arterial to equilibrium phases as critical for early HCC classification.
Conclusions:
- The novel GA feature selection approach effectively improves early HCC classification accuracy.
- Intensity features during specific contrast phases are highly informative for early HCC detection.
- This method offers a promising tool for enhancing early HCC diagnosis using MDCT.
