A Comparative Study on the Potential of Unsupervised Deep Learning-based Feature Selection in Radiomics

Area of Science:

• Radiomics and Medical Image Analysis
• Machine Learning Applications in Healthcare

Background:

• Deep learning systems are increasingly used in medical image analysis for radiomics.
• Feature-based systems remain preferred by physicians due to better explainability.
• High-dimensional data and low sample sizes present challenges like overfitting in machine learning.

Purpose of the Study:

• To investigate unsupervised deep learning-based methods for feature selection in radiomics.
• To compare the applicability and efficiency of five recent unsupervised deep learning algorithms.
• To evaluate the impact of these methods on classification results across diverse datasets.

Main Methods:

• Implementation and comparison of five novel unsupervised deep learning algorithms for feature selection.
• Testing on seven distinct datasets across three different application scenarios.
• Evaluation of feature selection performance and its effect on downstream classification tasks.

Main Results:

• Unsupervised deep learning-based feature selection significantly improved classification results.
• Performance gains were particularly notable when using small feature subsets.
• Deep learning methods outperformed conventional feature selection approaches.

Conclusions:

• Unsupervised deep learning offers an effective approach to feature selection in radiomics.
• This technique can identify and rank important features without requiring outcome data, aiding biomarker discovery.
• It holds significant potential for multiparametric radiology, enabling the identification of new biomarkers for treatment and prevention.