Spatial-channel attention-based stochastic neighboring embedding pooling and long-short-term memory for lung nodules classification

Handling lesion size and location variance in lung nodules are one of the main shortcomings of traditional convolutional neural networks (CNNs). The pooling layer within CNNs reduces the resolution of the feature maps causing small local details loss that needs processing by the following layers. In this article, we proposed a new pooling-based stochastic neighboring embedding method (SNE-pooling) that is able to handle the long-range dependencies property of the lung nodules. Further, an attention-based SNE-pooling model is proposed that could perform spatial and channel attention. The experimental results conducted on LIDC and LUNGx datasets show that the attention-based SNE-pooling model significantly improves the performance for the state of the art.