A Support Tensor Machine Based Algorithm for Distant Failure Prediction in Lung SBRT

Abstract

To develop a support tensor machine (STM) based model for predicting distant failure in early stage non-small cell lung cancer (NSCLC) treated with stereotactic body radiation therapy (SBRT) using pre-treatment PET and CT. The patient cohort used in this study includes 48 early stage NSCLC patients treated with SBRT from 2006 to 2012. Twelve patients (25%) failed at distant sites. For each patient, primary tumors are segmented on both pre-treatment CT and PET. Three-dimensional (3D) tensors of tumor for each imaging modality are constructed and used as the input for STM-based classifier. Different from the conventional radiomics approach that relies on handcrafted imaging features, the STM-based classifier uses 3D imaging as the input for the model. Thus it can take full advantage of the information from whole image without using handcrafted features. Compared with the conventional learning methods that are restricted to one-dimensional vectors as the input for modeling, the STM-based classifier can avoid the curse of dimensionality and small sample size problem since it uses 3D imaging not the vectorization imaging as the input. A STM iterative algorithm is used to train the weight vectors for every mode of the tensor. A 10-fold cross validation strategy is employed for the performance evaluation. The STM-based predictive algorithm is compared to support vector machine (SVM) based method where the vectorization of tumor image is used as the input. Two conventional radiomics approaches that use 29 handcrafted imaging features as the input are also used for comparison: 1) SVM-based classifier coupled with sequential feature selection (SVM-SFS); and 2) Logistic regression coupled with SFS (LR-SFS). Table I summarizes the performance of different predictive models, where area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN) and specificity (SPE) are utilized as the evaluation criteria. For both PET and CT, the STM-based predictive algorithm achieves highest AUC, ACC, SEN and SPE among four methods investigated in this study.Abstract 3638; Table IPerformance of different predictive methodsPETCTMethodAUCACCSENSPEAUCACCSENSPESVM0.76±0.0469.75±3.6168.33±13.6970.56±8.650.78±0.0572.92±6.0768.33±16.0374.44±11.01LR-SFS0.71±0.0369.58±3.4963.33±4.5671.67±4.560.72±0.0367.50±1.1463.33±4.5668.89±2.32SVM-SFS0.78±0.0168.75±4.5271.67±11.1864.11±9.500.76±0.0673.75±6.3573.33±7.4573.89±5.41STM0.85±0.0378.75±4.0680.00±6.2178.33±5.410.82±0.0376.75±3.7376.67±7.4576.77±4.65 Open table in a new tab We have developed a STM-based predictive algorithm using 3D image with an intrinsic 3D tensor structure as the input of models to predict distant failure of early stage NSCLC treated with SBRT. This STM-based method outperforms SVM with vectorization image data as the input, LR-SFS and SVM-SFS with handcraft features as the input.