Encrypted login | home

Program Information

Predict Clinical Outcome Through Quantitative Imaging Features Via Neural Network Models

no image available
x pan

X Pan*, X Qi , UCLA School of Medicine, Los Angeles, CA


SU-H4-GePD-J(B)-4 (Sunday, July 30, 2017) 4:30 PM - 5:00 PM Room: Joint Imaging-Therapy ePoster Lounge - B

Purpose: To apply neural networks to analyse quantitative imaging features to identify potential prognostic imaging features that are most relevant with treatment outcome.

Methods: A group of 59 head-and-neck patients with primary cancers in base-of-tongue area and their corresponding survival interval were collected and analysed in a single institution. Large number of quantitative image features of each patient was extracted based on a planning CT images, resulting in high-dimensional data. 80% patients were used to train the models, and remaining 20% were used for evaluation purposes. Principal component analysis was used to remove redundant imaging features before the imaging features fed into neural network. Two neural network models, including Evolutionary optimization neural network (EONN) and convolutional neural network (CNN), were developed to explore the association between the large number of radiomic features and RT clinical outcome. The EONN, adapted from back propagation neural network, was constructed based on the optimized initial weights. The initial weights were first optimized by evolutionary algorithm and then fed into EONN training model. The population was encoded in real number, the error between prediction and expected data is used as the fitness function.

Results: Better agreement was observed in EONN compared to the CNN model. The EONN model, based on 5 independent training experiments, yielded an averaged predicted survival interval (and standard deviation) of 30.3±0.2 months; the averaged predicted survival interval was 23.5±1.5 months in CNN model, compared to the reported value of 31.8 months for the test cohort.

Conclusion: We demonstrated the usage of the neural network models in RT outcome assessment. The EONN resulted in decent prediction of patient survival for high dimensional features in small sample in the question. The CNN, usually requires more connection between neurons, achieved an inferior training result given a relatively small data sample involving high dimensional features.

Contact Email: