Despite the success of Deep-Learning-based (DL) methods on Salient Object Detection (SOD), the need for abundantly labeled data and the high complexity of the network architectures limit their applications. Feature Learning from Image Markers (FLIM) is a recent methodology to build convolutional encoders with minimal human effort in data annotation. More recently, a FLIM encoder has been combined with an adaptive decoder to build flyweight FLIM networks for SOD, requiring only user-drawn markers in discriminative regions of a few (e.g., 4) images to train the entire model with no backpropagation. Furthermore, due to the data scarcity in some applications, using Cellular Automata (CA) may help compute better saliency maps. However, the initialization of the CA could be a problem since it is based on user input, priors, or randomness. Here, we propose a new strategy for CA initialization via a FLIM-based SOD network. In summary, CA interprets pixels of an initial saliency map as cells and cleverly designs transition rules to generate an improved saliency map through the evolution and interaction of each cell and its neighbors using the original pixel properties. CA requires initializing the cell's states, where methods diverge. By exploring the saliency map of a FLIM network, we circumvent the CA initialization problem and improve FLIM saliencies. Experiments in two challenging medical datasets demonstrate improvements in FLIM-based SOD, with results comparable to two state-of-the-art DL methods fine-tuned under data scarcity.