Abstract
In mammography, masses are primary indications of breast cancer, and it is necessary to classify them as malignant or benign. In this classification task, Computer Aided Diagnostic (CAD) systems using ensemble learning are able to assist radiologists in achieving better diagnoses of mammogram images. Ensemble learning consists of two steps: generating multiple base classifiers and then combining them together. However, combining all base classifiers in the ensemble model increases the computational cost and time. Therefore, ensemble pruning is an important step in ensemble learning to select the ensemble’s members. Due to the huge number of possible combinations in the ensemble, selecting the proper subset is a significant challenge. The objective of this paper is to select the optimal ensemble subset by using the Bee Algorithm (BA). A pool of different classifier models, such as Support Vector Machine (SVM), k-Nearest Neighbor (KNN), and Linear Discriminant Analysis (LDA), were trained using different samples of training data and 12 groups of features. Then, using this pool, BA was used to exploit and explore random uniform combination subsets of the trained classifiers. The mammogram image dataset used to test the model was collected from Hospital Kuala Lumpur (HKL) and consists of 263 benign and malignant masses. The proposed method achieves 77% Area Under Curve (AUC), 83% accuracy, 93% specificity, and 60% sensitivity.
Introduction and Problem Statement
Breast cancer remains a significant health challenge, and early detection through mammography is critical. Computer-Aided Diagnostic (CAD) systems have become vital tools for radiologists, often employing ensemble learning to improve predictive performance. Ensemble learning operates on the principle that a group of models can provide a more robust decision than any single model alone.
However, ensembles face a "Goldilocks" problem: too few models may miss subtle patterns, while too many models lead to high computational costs and potential performance degradation due to redundant or conflicting information. The challenge lies in ensemble pruning—efficiently searching the massive space of possible model combinations (e.g., $2^{50}$ for 50 models) to find the most accurate and diverse subset.
Methodology
The authors propose a three-phase framework centered on heterogeneity and nature-inspired optimization.
1. Pool Generation
To ensure a robust starting point, a diverse pool of base classifiers was created. This involved training SVM, KNN, and LDA models on various subsamples of the training data and 12 distinct groups of image features, including texture, shape, and frequency characteristics.
2. Ensemble Pruning via Bee Algorithm (BA)
The core innovation is the application of the Bee Algorithm to the pruning process. BA mimics the foraging behavior of honeybees:
- Scout Bees: Explore the search space of model combinations to identify promising subsets.
- Recruited Bees: Perform local searches around these promising subsets to refine the ensemble.
3. Fitness Function
The "quality" of a model subset was determined by a fitness function balancing Accuracy (measured by AUC) and Diversity. Diversity metrics, such as the Q-statistic and correlation coefficient, were used to ensure the selected models were complementary—meaning they do not all fail on the same types of data.
Experimental Results
The study utilized a dataset from Hospital Kuala Lumpur (HKL) containing 263 ROI-cropped mammogram images. The BA-pruned ensemble demonstrated strong performance:
- Accuracy: 83%
- Specificity: 93% (indicating high reliability in identifying benign masses)
- AUC: 77%
- Sensitivity: 60% (identifying malignant masses)
Statistically, the Bee Algorithm's superiority over traditional Genetic Algorithms (GA) was significant, with a p-value of < 0.00037. The results confirm that a pruned, heterogeneous ensemble can outperform the average performance of its individual members while reducing computational overhead.
Conclusion and Future Work
This research highlights that heterogeneity and diversity are more critical than the raw power of individual models in ensemble learning. While the system shows high specificity, the sensitivity of 60% suggests room for improvement, as missing a malignant mass is a critical risk in clinical settings. Future work will focus on expanding the dataset to larger public repositories like DDSM, penalizing false negatives more heavily in the fitness function, and integrating deep learning features into the heterogeneous pool.
Summary updated