Overview of the model architecture. The input image is presented on the left. The CNN has been divided into two components to enhance readability. The convolutional encoder (green) translates the image into a compressed features vector. Then, the convolutional decoder (blue) converts this vector into the final segmentation layers, which can be compared with the ones annotated by the experts. The rule-based classifier is depicted on the right. Given the predicted segmentation layers, this module casts them to one of the 4 classes of interest.