Quantitative Imaging Features Predict Clinical Survival: An Initial Analysis with Neural Network Models

Abstract

Tumor imaging features may predict treatment outcome and guide individualized treatment in radiotherapy (RT). However, given high dimensional imaging features versus relatively small patient sample, identifying potential prognostic imaging biomarkers is typically challenging. The aim of the work is to identify prognostic imaging features that are most relevant with patient RT outcome (i.e., survival interval) via neural network models. Evolutionary optimization neural network (EONN), based on back propagation neural network (BPNN), was developed to explore the association between the large number of radiomic features and RT clinical outcome. The advanced convolutional neural network (CNN) was also used to cross check the findings. The EONN is constructed based on the optimized initial weights of the BPNN. Evolutionary algorithm (EA) is used to optimize the initial weights and thresholds fed into training model. The population of EA is encoded in real number, and the error between the prediction and the expected data is used as the fitness function. To evaluate the efficacy of the models, a group of 59 base-of-tongue patients and corresponding survival interval were analyzed, where 47 patients (80%) were used as a training cohort, and remaining 12 patients (20%) were used as a test cohort. For each patient, a large number of quantitative image features (including shape, texture, intensity, etc.) were extracted based on a pre-treatment CT images, resulting in 1387-dimensional data in six categories. Principal component analysis was used to remove redundant imaging features before the features fed into neural network. Table 1 summarizes the predicted survival results using neural networks. EONN resulted in the best performance among three models. The BPNN, based on simple training method, yielded the largest deviations. The CNN, usually requires more connection between neurons, didn’t achieve a good training result when a relatively small data sample involving high dimensional features. We demonstrated the application of the neural network models in RT radiomic analysis. The EONN resulted in decent prediction results of patient survival for high dimensional features in small sample. The identified prognostic image features can be adopted to design individualized radiation treatment to potentially improve clinical outcomes.Abstract 1031; Table 1Prediction results using three neural network modelsBPNNEONNCNNTest pt #Survival interval (Mon.)Predicted value (Mon.)DeviationPredicted value (Mon.)DeviationPredicted value (Mons)Deviation19.355.85.012.50.3433.6215.527.20.822.50.4130.2312.223.20.916.20.3271.243.431.98.53.10.130.18.0515.513.60.113.10.2140.1672.447.20.366.50.160.90.2775.522.60.776.50.026.90.68110.6-34.41.3930.22.91.092055.81.818.80.117.40.11015.9-17.32.17.20.5-7.81.51110.92.30.810.80.020.80.91221.1280.323.50.133.20.6Mean31.821.31.930.30.223.51.5 Open table in a new tab