Multi-Resolution based Deep Learning and Knowledge Transfer for Pulmonary Nodule Candidate Identification

Abstract

Nowadays, the classification of lung nodules is an essential to identify and diagnose the lung tumors at its prior stage. But, this essential process was affected because of different sizes and shapes of lung nodules. To tackle these problems, Multi-Resolution Convolutional Neural Network with Knowledge Transfer (MRCNN-KT) was suggested for extracting the characteristics of manylevels and resolutions from various depth layers in the CNN to categorize the lung nodule candidates. In contrast, this 2D-CNN framework was restricted in extracting the contextual details between multiple slices. Therefore this article proposes an MR 3D-CNN with KT (MR3DCNN-KT) framework in which 3D convolutions are applied for extracting the features from both spatial and temporal dimensions between slices. Therefore, the contextualdetails encoded in the multiple adjacent slices is obtained.Also, multiple channels of data are created from the input frames and the data from each channel is fused to get the final feature. Moreover, a regularization and fusion methods are proposed forfurther improving the classification efficiency. Finally, the effectiveness of the MR3DCNN-KT framework is exhibited and evaluated with the MRCNN-KT via the experimental results.