(Tuesday, 7/14/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room: Track 2
Purpose: Novel actuarial deep learning neural networks (ADNNs) are proposed for joint prediction of radiotherapy outcomes, e.g., radiation pneumonitis (RP) and local control (LC) in stage III non-small cell lung cancer (NSCLC) patients. Unlike traditional TCP/NTCP models that only use dosimetric information, our models consider complex interactions among multi-omics information; i.e., PET radiomics, cytokines and micro-RNA. Additionally, time-to-event information was utilized in the actuarial prediction.
Methods: Two models were investigated: ADNN-DVH considered dosimetric information only, ADNN-com additionally considered imaging and biological information. In both models, differential dose volume histograms (DVHs) were fed into 1D convolutional neural networks (CNNs) for dimensionality reduction. Variational encoders (VAEs) were used in ADNN-com to learn representation of imaging and biological data. Reduced representations were finally fed into Surv-nets to predict event-free probabilities in different time intervals. Accordingly, time-to-event information was incorporated into designated loss functions.
Results: Models were evaluated on 117 patients and were independently tested on 25 newly treated patients. A multi-institutional RTOG0617 dataset was also used for external validation. ADNN-DVH yielded a cross-validated c-indexes [95% Cis] of 0.660 [0.630 -0.690] for RP2 prediction and 0.727 [0.700-0.753] for LC prediction, outperforming a generalized Lyman model for RP2 (0.613 [0.583-0.643]) and a generalized log-logistic model for LC (0.569 [0.545-0.594]). ADNN-com achieved an even better performance with c-indexes of 0.705 [0.676-0.734] for RP2 and 0.740 [0.714-0.765] for LC. In independent testing, Lyman model, ADNN-DVH and ADNN-com showed c-indexes of 0.588, 0.667, 0.691, respectively for RP2, and 0.573, 0.706 and 0.721, respectively for LC. In external testing, ADNN-DVH showed indexes of 0.762 for RP and 0.618 for LC, outperforming a Lyman model for RP2 (c-index 0.736) and log-logistic model for LC (c-index 0.554).
Conclusion: Novel deep learning architectures that integrates multi-omics information outperformed traditional TCP/NTCP models in actuarial prediction of LC and RP2.
Keywords
Not Applicable / None Entered.
Taxonomy
TH- Response Assessment: Modeling: Machine Learning
Contact Email
