Deep Learning- and Expert Knowledge-Based Feature Extraction and Performance Evaluation in Breast Histopathology...

Area of Science:

• Oncology
• Medical Imaging
• Computer Science

Background:

• Cancer diagnosis relies on analyzing histopathology images to detect abnormal cell growth.
• Computer-aided detection (CAD) systems using knowledge-based or deep learning approaches are crucial for enhancing diagnostic accuracy.
• Effective feature extraction from histopathology images is essential for the performance of CAD systems.

Purpose of the Study:

• To evaluate and compare the performance of three distinct feature extraction methods for breast cancer diagnosis.
• To assess the efficacy of Convolutional Neural Networks (CNNs), VGG16 transfer learning, and a knowledge-based system in identifying cancerous tissues.
• To determine the optimal feature extraction technique when combined with various classification algorithms.

Main Methods:

• Three feature extraction techniques were employed: a Convolutional Neural Network (CNN), the VGG16 transfer learning architecture, and a knowledge-based system.
• Extracted features were tested using seven different classifiers: Neural Network (64 units), Random Forest, Multilayer Perceptron, Decision Tree, Support Vector Machines, K-Nearest Neighbors, and Narrow Neural Network (10 units).
• The BreakHis 400× histopathology image dataset was utilized for performance evaluation.

Main Results:

• The CNN method achieved a maximum accuracy of 85% when paired with Neural Network and Random Forest classifiers.
• The VGG16 transfer learning approach reached up to 86% accuracy with the Neural Network classifier.
• The knowledge-based feature extraction method demonstrated superior performance, achieving up to 98% accuracy with Neural Network, Random Forest, and Multilayer Perceptron classifiers.

Conclusions:

• Knowledge-based feature extraction significantly outperforms deep learning methods (CNN, VGG16) in breast cancer diagnosis using the BreakHis dataset.
• The combination of knowledge-based features and classifiers like Neural Network, Random Forest, or Multilayer Perceptron offers a highly accurate approach for histopathology-based cancer detection.
• Further research into knowledge-based systems could lead to more robust and accurate computer-aided diagnosis tools for oncology.