Incorporating Deep Learning Techniques into Outcome Modeling in Non-Small Cell Lung Cancer Patients after Radiation Therapy

Abstract

Radiation therapy (radiotherapy) together with surgery, chemotherapy, and immunotherapy are common modalities in cancer treatment. In radiotherapy, patients are given high doses of ionizing radiation which is aimed at killing cancer cells and shrinking tumors. Conventional radiotherapy usually gives a standard prescription to all the patients, however, as patients are likely to have heterogeneous responses to the treatment due to multiple prognostic factors, personalization of radiotherapy treatment is desirable. Outcome models can serve as clinical decision-making support tools in the personalized treatment, helping evaluate patients’ treatment options before the treatment or during fractionated treatment. It can further provide insights into designing of new clinical protocols. In the outcome modeling, two indices including tumor control probability (TCP) and normal tissue complication probability (NTCP) are usually investigated. Current outcome models, e.g., analytical models and data-driven models, either fail to take into account complex interactions between physical and biological variables or require complicated feature selection procedures. Therefore, in our studies, deep learning (DL) techniques are incorporated into outcome modeling for prediction of local control (LC), which is TCP in our case, and radiation pneumonitis (RP), which is NTCP in our case, in non-small-cell lung cancer (NSCLC) patients after radiotherapy. These techniques can improve the prediction performance of outcomes and simplify model development procedures. Additionally, longitudinal data association, actuarial prediction, and multi-endpoints prediction are considered in our models. These were carried out in 3 consecutive studies. In the first study, a composite architecture consisting of variational auto-encoder (VAE) and multi-layer perceptron (MLP) was investigated and applied to RP prediction. The architecture enabled the simultaneous dimensionality reduction and prediction. The novel VAE-MLP joint architecture with area under receiver operative characteristics (ROC) curve (AUC) [95% CIs] 0.781 [0.737-0.808] outperformed a strategy which involves separate VAEs and classifiers (AUC 0.624 [ 0.577-0.658]). In the second study, composite architectures consisted of 1D convolutional layer/ locally-connected layer and MLP that took into account longitudinal associations were applied to predict LC. Composite architectures convolutional neural network (CNN)-MLP that can model both longitudinal and non-longitudinal data yielded an AUC 0.832 [ 0.807-0.841]. While plain MLP only yielded an AUC 0.785 [CI: 0.752-0.792] in LC control prediction. In the third study, rather than binary classification, time-to-event information was also incorporated for actuarial prediction. DL architectures ADNN-DVH which consider dosimetric information, ADNN-com which further combined biological and imaging data, and ADNN-com-joint which realized multi-endpoints prediction were investigated. Analytical models were also conducted for comparison purposes. Among all the models, ADNN-com-joint performed the best, yielding c-indexes of 0.705 [0.676-0.734] for RP2, 0.740 [0.714-0.765] for LC and an AU-FROC 0.720 [0.671-0.801] for joint prediction. The performance of proposed models was also tested on a cohort of newly-treated patients and multi-institutional RTOG0617 datasets. These studies taken together indicate that DL techniques can be utilized to improve the performance of outcome models and potentially provide guidance to physicians during decision making. Specifically, a VAE-MLP joint architectures can realize simultaneous dimensionality reduction and prediction, boosting the performance of conventional outcome models. A 1D CNN-MLP joint architecture can utilize temporal-associated variables generated during the span of radiotherapy. A DL model ADNN-com-joint can realize multi-endpoint prediction, which allows considering competing risk factors. All of those contribute to a step toward enabling outcome models as real clinical decision support tools.