Purpose: To design and develop a novel composite deep learning architecture incorporating serial PET imaging, biological and dosimetric data for joint actuarial predictions of lung pneumonitis (RP) and local control (LC) of non-small cell lung cancer (NSCLC) patients after radiotherapy.
Methods: Data were obtained from 88 NSCLC patients, 27 of whom failed locally after radiotherapy, 18 of whom had RP grade â‰¥ 2 (RP2). EQD2 dose volume histograms (DVHs) for GTV and lung-GTV, cytokines levels, micro-RNA expressions, single-nucleotide polymorphism (SNP) and texture radiomics extracted from PET (GTV) were collected as inputs for the proposed architecture. The deep learning architecture comprises variational autoencoders (VAEs) aimed for representation learning and multi-layer perceptrons (MLPs) designated for prediction. Specifically, 4 VAEs were implemented for GTV and lung â€“ GTV DVHs, biological data and PET imaging data, respectively. Latent variables of GTV DVH and PET tumor radiomics, together with 2 out of 8 latent variables of biological data were fed into the MLP for LC prediction. Latent variables of lung-GTV DVH and 6 out of 8 latent variables of biological data were used for RP2 prediction. A prediction loss defined as a negative loglikelihood function incorporating discrete time-to-event information was considered instead of commonly used ones in MLP classification. The performance was evaluated using cross-validation and assessed by the receiver-operating characteristics area under the curve (AUC).
Results: The joint actuarial-analysis architecture achieved an AUC of 0.702 on RP2 and 0.712 on LC, which outperformed the prevalent Lyman model of RP2 (AUC: 0.547) and empirical log-logistic model of LC (AUC: 0.615).
Conclusion: The feasibility of deep architecture for joint actuarial prediction of LC and RP2 in NSCLC has been demonstrated. Specifically, the joint prediction of the deep architecture outperformed the prevalent models in LC and PR2 prediction, respectively.