Radiomics-Based Model Using Tumor and Peritumoral Features with Semi-Supervised and Privileged Learning for...

Lung cancer is the leading cause of cancer-related mortality globally. Early detection of high-risk patients for local or distant metastasis is challenging for better monitoring and treatment planning. Machine learning models have been proposed for diagnosis and prediction of metastasis risk. However, limited data, due to long follow-up period, and lack of multi center cohorts affect model generalization as they overlook the impact of various data heterogeneity.

The aim of this study is to develop a generalizable radiomics-based model, capable of predicting the occurrence of metastasis within two years of the initial lung cancer diagnosis combining supervised techniques with semi-supervised and privileged learning approaches.

Data from 114 lung cancer patients, without metastasis during diagnosis, were used for model training. Computed tomography (CT) images, acquired from different vendor machines, segmentation of the tumor and clinical data were available for each patient. Radiomic features were extracted from the primary tumor as well as from an area reflecting the tumor rim, covering both peri- and extra-tumoral regions of diverse widths. Feature harmonization was applied among the features from different CT scanners. A Support Vector Machines (SVM) model with hyperparameter tuning was trained using the most important radiomics and clinical features that were selected by iterative recursive feature elimination. Semi-supervised learning was used to increase the training dataset, where metastasis labels were missing. Finally, privileged learning was applied to train the SVM using additional clinical data that may not be available in the testing dataset. All models were evaluated using 21 patients from two clinical sites. Human derived predictions were compared with the AI-based ones.

The width of the tumor rim that maximized the balanced accuracy (BA) during the training of the SVM model extended 3mm inside and 3mm outside the tumor edge. The selected radiomic features had higher values in patients with metastasis, suggesting a higher heterogeneity of the analysis areas. The SVM model had a BA of 76.7 % with high specificity and low sensitivity (86.7 % and 66.7 %, respectively) in the external dataset. The application of semi-supervised learning improved performance, while the inclusion of privileged learning techniques led to a BA of 81.65 % improving both specificity and sensitivity (80 % and 83.3 % respectively). This model outperformed the human expert in terms of BA for data, in a small dataset, from one clinical site.

The rim size analysis provided insights into tumor aggressiveness and metastasis occurrence, especially when combined with the patient's stage and tumor characteristics. Integrating semi-supervised and privileged learning techniques significantly enhanced SVM models' performance in identifying lung cancer patients at risk for metastasis, outperforming human-derived predictions in a small external dataset. The model's generalizability was demonstrated on an external validation set from two different regions. More extensive validation is necessary to confirm these findings.