Non-linear Logistic Regression applied to Radiomics.

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Radiomics offers significant potential for cancer detection and treatment response prediction by analyzing tumor phenotypic characteristics.
• Despite its promise, radiomics is not yet widely adopted in clinical practice.
• Reliable machine learning (ML) approaches are crucial for advancing radiomics utilization.

Purpose of the Study:

• To improve the performance of Logistic Regression (LR), a commonly used ML model in radiomics.
• To explore the application of one-hot encoding for representing radiomics features.
• To demonstrate the effectiveness of a proposed non-linear LR model in radiomics datasets.

Main Methods:

• Radiomics features were represented using one-hot encoding.
• A non-linear Logistic Regression (LR) model was developed based on this representation.
• The performance of the proposed LR model was evaluated on two radiomics datasets.

Main Results:

• The proposed one-hot encoding transforms the LR model into a non-linear classifier.
• The resulting non-linear LR demonstrated performance comparable to the Naive Bayes classifier (NBC).
• The additive nature of the LR score function allows for easy measurement of individual feature contributions.

Conclusions:

• The developed non-linear LR model, utilizing one-hot encoding, shows promise for enhancing radiomics analysis.
• This approach could facilitate the clinical integration of radiomics by providing interpretable and effective ML models.
• Further validation on diverse radiomics datasets is warranted to confirm clinical utility.