Medical image segmentation is a pertinent issue, with deep learning being a leading solution. However, the demand for a substantial number of fully annotated images for training extensive models can be a hurdle, particularly for applications with diverse images, such as brain tumors, which can manifest in various sizes and shapes. In contrast, the recent Feature Learning from Image Markers (FLIM) methodology, which incorporates an expert in the learning process, has proven to be effective. This approach generates compact networks requiring only a few images to train the convolutional layers without backpropagation. In our study, we implement an interactive method for image selection and training based on FLIM, exploring the user's knowledge. The results underscore the efficacy of our methodology, as we were able to select a small set of images to train the encoder of a U-shaped network, achieving performance on par with manual selection and even surpassing the same U-shaped network trained by backpropagation with all training images.