Absolute cosine-based SVM-RFE feature selection method for prostate histopathological grading.

Area of Science:

• Histopathology
• Computational Pathology
• Medical Image Analysis

Background:

• Feature selection (FS) is crucial for grading and diagnosing prostate histopathological images.
• High-dimensional texture features from tissue components (lumen, nuclei, cytoplasm, stroma) are challenging to represent.
• Overfitting is a significant issue due to the complexity of these features.

Purpose of the Study:

• To develop a novel FS method for selecting minimally redundant features from histopathological images.
• To improve the accuracy of classifying prostate and colon cancer based on tissue component textures.
• To address the challenge of high-dimensional texture representation in medical image analysis.

Main Methods:

• Proposed a new FS method, Support Vector Machine-Recursive Feature Elimination with Absolute Cosine (SVM-RFE(AC)).
• Integrated SVM-RFE with an absolute cosine (AC) filter to reduce feature redundancy.
• Utilized single and ensemble classification models, including a product rule merging method.

Main Results:

• Experiments were conducted on H&E stained prostate and colon cancer images.
• The SVM-RFE(AC) method demonstrated superiority over other SVM and SVM-RFE-based methods.
• Accurate classification was achieved for prostate cancer grades (benign vs. grade 3, benign vs. grade 4, grade 3 vs. grade 4) and colon cancer grades (1 vs. 2).

Conclusions:

• The developed SVM-RFE(AC) method successfully identifies critical texture features in tissue components.
• This method enables accurate distinction between various Gleason grades in prostate cancer.
• The proposed FS approach significantly outperforms existing methods in histopathological image analysis.