Multi-Objective-Based Radiomic Feature Selection for Lesion Malignancy Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate classification of lesion malignancy is crucial for reducing false positives in medical screening.
• Radiomics offers potential for differentiating malignant from benign tumors using quantitative image features.
• Selecting optimal radiomic features is critical for effective classification models.

Purpose of the Study:

• To develop a novel multi-objective feature selection (MO-FS) algorithm for improved malignancy classification.
• To simultaneously consider sensitivity and specificity as objective functions in feature selection.
• To enhance the effectiveness of radiomic feature selection strategies.

Main Methods:

• Proposed a multi-objective based feature selection (MO-FS) algorithm.
• Developed a modified entropy-based termination criterion for automatic algorithm stopping.
• Designed a solution selection methodology for multi-objective learning (SMOLER) using evidential reasoning.
• Implemented an adaptive mutation operation for automatic mutation probability generation.

Main Results:

• Evaluated MO-FS for classifying lung nodule malignancy in low-dose CT scans.
• Assessed MO-FS for classifying breast lesion malignancy in digital breast tomosynthesis.
• Demonstrated superior classification performance of MO-FS selected features compared to other methods.

Conclusions:

• The proposed MO-FS algorithm achieves better classification performance.
• MO-FS provides a general and effective strategy for radiomic feature selection.
• This approach can enhance diagnostic accuracy in medical imaging.