Predicting Lymph Node Metastasis of Lung Cancer Using Stacked Sparse Autoencoder

Predicting accurately the lymph node metastasis is very essential for right staging and making the appropriate treatment strategies for lung cancer. Existing prediction methods are mostly based on the image features, which depend greatly on the postprocessing algorithms. To deal with this problem, we proposed to use the stacked sparse autoencoder (SSAE) network to predict the lymph node metastasis in lung cancer. To assess the performance of SSAE, the radiomics features based prediction and image based SSAE prediction were quantitatively compared in terms of ROC and AUC. In the radiomics-based prediction, more than 400 features were fed into a SSAE network, and in the image-based SSAE prediction, 100 images were inputted. The experimental results show that the image-based SSAE experienced a better performance than that radiomics-based one, even with a fewer samples to train the network. The AUC obtained by image-based SSAE was enhanced by about 50%.