Prediction of Prognosis for Nonsmall Cell Lung Cancer after Targeted Therapy Based on CT Imaging and Dense Convolutional Sparse Coding

The prediction of prognosis for epidermal growth factor receptor (EGFR) mutation-driven nonsmall cell lung cancer (NSCLC) after targeted therapy has attracted a lot of attention. Traditional machine learning (e.g., random forest) is with limited capability to handle complex CT features without preprocessing. CNN is a commonly used deep learning architecture; however, the architecture lacks a certain degree of physical interpretability and requires a large number of network parameters. In this study, a dense convolutional sparse coding (DCSC) model is proposed to improve prediction accuracy. The constructed model is mainly composed of convolutional sparse coding (CSC), dense connection, and attention mechanism. Operation procedure for predicting prognosis is as follows. First, tumor segmentation is performed manually using ITK-SNAP software to obtain tumor masks for CT images. Subsequently, CT images and corresponding masks are input into pyradiomics to extract tumor features. Finally, extracted features are input into the proposed DCSC model for predicting prognosis [progressive disease (PD) or partial response (PR)]. The proposed DCSC model exhibits high accuracy and stability with two datasets including 101 and 388 CT images from NSCLC patients. Area under ROC curve (AUC) ranges from 0.79 to 0.88 and accuracy ranges from 0.8 to 0.85. Further quantitative comparison indicates that the proposed DCSC model outperforms traditional CNN and random forest.