Integration of Radiomics and Tumor Biomarkers in Interpretable Machine Learning Models.

Area of Science:

• Oncology
• Radiology
• Machine Learning

Background:

• Deep neural networks (DNNs) show promise in medical imaging but lack interpretability, hindering clinical adoption for cancer diagnosis.
• Interpretability is crucial for clinicians to trust and utilize AI-driven diagnostic tools in oncology.

Purpose of the Study:

• To develop an interpretable machine learning model, ConRad, for lung cancer classification using CT scans.
• To integrate expert-derived radiomics and DNN-predicted biomarkers into a transparent classifier.

Main Methods:

• Proposed ConRad, combining concept bottleneck model (CBM)-predicted biomarkers and radiomics features in interpretable classifiers.
• Evaluated ConRad against black-box convolutional neural networks (CNNs) using segmented CT scans.
• Assessed combinations of radiomics, predicted biomarkers, and CNN features across five classifiers, utilizing Lasso for feature selection.

Main Results:

• ConRad models, particularly nonlinear SVM and logistic regression with Lasso, demonstrated superior performance in five-fold cross-validation.
• Interpretability was a key advantage of ConRad over traditional CNNs.
• Lasso feature selection reduced model complexity and improved accuracy.

Conclusions:

• The ConRad model effectively combines CBM-derived biomarkers and radiomics for interpretable lung nodule malignancy classification.
• ConRad offers a promising, transparent alternative to black-box models in oncological imaging analysis.
• The approach enhances clinical trust and utility of AI in cancer diagnosis.