Enhanced Lung Cancer Prediction via Integrated Multi-Space Feature Adaptation, Collaborative Alignment and Disentanglement Learning

Abstract

Lung cancer, marked by the rapid and uncontrolled proliferation of abnormal cells in the lungs, continues to be one of the leading causes of cancer-related deaths globally. Early and accurate diagnosis is critical for improving patient outcomes. This research presents an enhanced lung cancer prediction model by integrating Adaptation Multiple Spaces Feature and L1-norm Regularization (AMSF-L1ELM) with Primitive Generation with Collaborative Relationship Alignment and Feature Disentanglement Learning (PADing). Initially, the AMSF-L1ELM model was employed to address the challenges of feature alignment and multi-domain adaptation, achieving a baseline performance with a test accuracy of 83.20%, precision of 83.43%, recall of 83.74%, and an F1-score of 83.07%. After incorporating PADing, the model exhibited significant improvements, increasing the test accuracy to 98.07%, precision to 98.11%, recall to 98.05%, F1-score to 98.06%, and achieving a ROC-AUC of 100%. Cross-validation results further validated the model's robustness, with an average precision of 99.73%, recall of 99.55%, F1-score of 99.64%, and accuracy of 99.64% across five folds. The study utilized four distinct datasets covering a range of imaging modalities and diagnostic labels: the Chest CT-Scan dataset from Kaggle, the NSCLC-Radiomics-Interobserver1 dataset from TCIA, the LungCT-Diagnosis dataset from TCIA, and the IQ-OTH/NCCD dataset from Kaggle. In total, 4085 images were selected, distributed between source and target domains. These results demonstrate the effectiveness of PADing in improving the model's performance and enhancing lung cancer prediction accuracy across multiple domains in complex medical imaging data.