Exploring Several Novel Strategies to Enhance Prediction of Lung Cancer Survival Time

This research evaluates several strategies to predict Lung cancer (LCa) survival: i) comparing PET vs. CT imaging; ii) handcrafted vs. deep radiomic features (HRF vs. DRF); iii) semi-supervised vs. supervised learning; iv) sets with clinical features (CF) vs. HRF alone and DRF alone; and v) 3 different masks-whole, cropped ($32 \times 32 \times 32 \mathrm{~mm}^{3}$), and segmented PET/CT images, vi) 3 regression algorithms (RA). Our analysis included 199 LCa patients (just 99 with existing death time). We extracted HRFs and DRFs by PySERA and a 3D Autoencoder within the ViERA software (visear.ca), applying them in both supervised and semi-supervised frameworks to predict LCa survival. In the supervised model, we used Principal Component Analysis (PCA) followed by 3 RAs, applied to various combinations of CFs, HRFs, and DRFs. Additionally, the semi-supervised model augmented our dataset by pseudo-labeling patients with missing outcomes. The models were evaluated via 5-fold cross-validation and external nested testing. This study showed that i) PET significantly outperformed CT images with an averaged mean absolute error (MAE) $\sim 0.57 \pm 0.27$ [outcome range 0.1 to 6.6 years] using semi-supervised approach+HRF extracted from PET+PCA+ Multilayer Perceptron Regressor (MLPR); ii) HRFs outperformed DRFs with averaged MAE~0.57 $\pm \mathbf{0 . 2 7}$ using above ML; iii) Semisupervised significantly outperformed supervise learning with averaged MAE~0.52±39 using sole CF the above ML; and iv) combinations with CFs outperformed combinations without CFs, with averaged MAE~0.52 ±0.39 using semi-supervised learning, sole CF and above ML; and v) whole mask outperformed others with averaged MAE~0.59 ±0.27 using semi-supervised learning+DRFs extracted from whole CT images +CF+ above ML, vi) MLPR outperformed other algorithms with a averaged MAE of $0.52 \pm 0.39$ using sole CFs and semi-supervised learning.