Radiation and Cancer Physics

SS 01 - Physics 1 -Best of Physics

7 - Prediction of Local Control in Non-Small Cell Lung Cancer Patients after Radiation Therapy By Composite Deep Learning Neural Networks

Sunday, October 21
1:25 PM - 1:35 PM
Location: Room 214 A/B

Prediction of Local Control in Non-Small Cell Lung Cancer Patients after Radiation Therapy By Composite Deep Learning Neural Networks
S. Cui1, Y. Luo2, S. Jolly2, R. K. Ten Haken2, and I. El Naqa2; 1University of Michigan, Ann Arbor, Ann Arbor, MI, 2Department of Radiation Oncology, University of Michigan, Ann Arbor, MI

Purpose/Objective(s): Despite improvements in systemic therapy and radiation techniques, local control (LC) in locally advanced non-small cell lung cancer (NSCLC) after definitive radiation therapy (RT) remains poor. Current models for LC prediction have had limited clinical applicability due to their poor performance under the circumstance of relatively small sample sizes. Small sample sizes have compromised the training of traditional machine learning methods including fully-connected artificial neural networks such as multi-layer perceptrons (MLP). Here, advanced composite architectures based on deep learning (DL) techniques that can take advantage of associations among longitudinal (sequential) personalized patient data to reduce data size requirements will be investigated for potential improved prediction of LC in NSCLC patients post-radio therapy.

Materials/Methods: Data were obtained under IRB approval from 99 NSCLC patients, 29 of whom failed locally after radiotherapy. Dosimetric, cytokines levels, single-nucleotide polymorphism (SNP), and texture radiomics extracted from PET images were collected and incorporated into DL architectures for prediction of LC. Among these variables, cytokines were collected at three time points and PET images were collected before and during the treatment. Two architectures were proposed to account for the temporal associations in these longitudinal data. The first applied 1D convolutional (1D-Conv) layers to the longitudinal data, then merged the outputs with the rest of the features and fed them into the subsequent fully-connected (FuCon) layers. The other architecture was similar to the first except locally-connected layers (LoCon) replaced 1D-Conv layers. An MLP model was also built for comparison with two proposed architectures. The predictive models were implemented on a high performance computing cluster using the python DL library Keras. The adaptive-learning-rate method, RMSprop was implemented for DL training and synthetic minority over-sampling (SMOTE) was applied to the training set during cross-validation (CV) to mitigate class imbalance problems. The performance was evaluated using CV and assessed by the receiver-operating characteristics area under the curve (AUC).

Results: The composite architecture of 1D-Conv layers and FuCon layers achieved an AUC of 0.83 (95%CI: 0.807-0.841). Whereas, the composite architecture of LoCon layers and FuCon layers achieved an AUC of 0.80 (95%CI 0.775-0.811). Both outperformed MLP which had an AUC of 0.78 (95%CI: 0.751-0.790);

Conclusion: The feasibility of applying deep learning methods for predicting LC in NSCLC has been demonstrated. Specifically, 1D-Conv layers performed the best in incorporating longitudinal data in the prediction model. Making use of temporal association by 1D-Conv layers significantly improved LC prediction compared to MLP. This work may be used prospectively to individualize radiation dosing or guide in altering systemic therapy in NSCLC patients.

Author Disclosure: S. Cui: None. S. Jolly: None. R.K. Ten Haken: Research Grant; NIH-NCI. Honoraria; University of Copenhagen. Travel Expenses; Varian Medical Systems Inc, University of Copenhagen. I. El Naqa: None.

Sunan Cui, BS

Disclosure:
No relationships to disclose.

Presentation(s):

Send Email for Sunan Cui


Assets

7 - Prediction of Local Control in Non-Small Cell Lung Cancer Patients after Radiation Therapy By Composite Deep Learning Neural Networks



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Prediction of Local Control in Non-Small Cell Lung Cancer Patients after Radiation Therapy By Composite Deep Learning Neural Networks