Lung Cancer Classification Using CapsNets

Convolution Neural Networks being computationally intense, can automatically detect feature maps from the images. Although the performance of CNN is excellent, there is a loss of valuable information when using pooling layers. The concept of CapsNets, which overcomes the loss of information during pooling operations, is promising in deep learning. The system provides subpar results on benchmark data sets that have complex setups. Congruent with the excellent results achieved by Convolutional Neural Networks (CNNs), namely DeepCaps, an architecture based on convolutional 3D models and a dynamic routing process. With Deep Caps, we significantly exceed the current state of the art in the CapsNets on LuNa-16, while achieving a significant percentage decrease in parameter count. In addition, we present a decoder network (independent of class) that enables us to apply loss reconstruction as a regularization term. Thus, we come to another fascinating attribute of the decoder, allowing us to manipulate the image’s physical representation determined by its instantiation parameters.