Figure 5.

The architecture of capsule networks along with the decoder module. (a) A simple CapsNet model achieves accuracy that is comparable to best CNN till now. The length of class instantiation activity vector of “Normal” and “Pneumonia” classes in output capsule layer indicates presence of one of the class (in a specific CXR image), which is used to calculate classification loss. As discussed above, weight matrix $M_{xy}$ is a matrix of weights between both ${\alpha }_x$, $x\in$ (1, 16 × 48 × 48) in Primary Caps layer and$v_y$, $y\in$ (1,2), where 1 represents “Normal” and 2 represents “Pneumonia”. (b) Decoder structure for reconstruction of CXR images from the class representation. The mean square error is minimized during the training phase of the network. The decoder has comparatively subservient role to the CapsNet in this research as it is used for encoding and regularizing the activity instantiation vector. In this research, the decoder will mostly be stuck in a local optimum due to less no. of capsules mapping values to the reconstruction layer, which leads to a lower accuracy of the image.