Title: Leveraging Deep Neural Network Gradient Loss for Radiomic Feature Selection in Lung Cancer Staging

Abstract: The extraction of quantitative imaging biomarkers from medical scans through radiomics has established itself as a vital component in computer-aided cancer diagnostics. Nevertheless, radiomics datasets often suffer from a high dimensionality relative to a scarcity of samples, rendering robust feature selection essential for developing dependable predictive models. To address this challenge, this research introduces a Gradient-Loss Recursive Feature Elimination (GL-RFE) framework. This approach utilizes gradient sensitivity analysis derived from deep neural networks to pinpoint the most significant radiomic features for determining lung cancer stages.

Using the PyRadiomics extension within the 3D Slicer platform, researchers extracted 106 radiomic features from chest Computed Tomography (CT) scans. The GL-RFE method assesses the importance of each feature by calculating the gradients of the network’s loss function concerning the input features, subsequently removing those that contribute least through a recursive process. The final selection of the top 15 features was employed to train a deep neural network classifier capable of differentiating between early-stage and advanced-stage lung cancer.

The GL-RFE framework demonstrated robust classification capabilities, yielding an accuracy of 90.22%, precision of 90.10%, recall of 90.24%, and an F1-score of 90.16% on the test dataset. Further validation through visualization techniques, such as distribution plots and correlation heat maps, highlighted a decrease in feature redundancy alongside enhanced separability between classes. When contrasted with traditional feature selection methods, GL-RFE proved superior in capturing nonlinear interactions among features and boosting the model’s generalization ability. This protocol offers a reproducible and interpretable methodology for radiomics-driven cancer staging, making it especially effective for high-dimensional, small-sample biomedical data, with potential utility extending to fields like genomics and multimodal clinical analysis.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC