Enhancing Transferability of Features from Pretrained Deep Neural Networks for Lung Nodule Classification

Among most popular feature extractors, pretrained deep neural networks play a central role in transfer learning to extract high-level feature on small datasets. The transferable performance, however, cannot be guaranteed for the task of interest. To enhance the transferability, this paper employs fine-tuning and feature selection in a different way to improve the accuracy of lung nodule classification. The fine-tuning technique retrains the neural network using lung nodule dataset, while feature selection captures a useful subset of features for lung nodule classification. Preliminary experimental results on CT images from Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) confirm that the classification accuracy on lung nodule can be significantly improved via finetuning and feature selection. Furthermore, the results outperform competitively handcrafted texture descriptors.