Interactive Ground-Truth-Free Image Selection for FLIM Segmentation Encoders

Deep Learning has shown impressive results in computer vision tasks at the cost of becoming increasingly extensive and requiring a voluminous amount of labeled images. Feature Learning from Image Markers (FLIM) is a method to build convolutional encoders with minimal user effort in image annotation. FLIM estimates the encoder's filters directly from markers drawn by the user (scribbles, clicks) in significant regions of a few representative images (e.g., 9). However, the selection of those few images significantly impacts the model's performance, and image selection is mainly addressed by either the user inspecting all images and subjectively selecting the most relevant ones or by using a method that requires pixel-wise annotation (ground-truth) for recommending images. The main goal of our work is to make this task less subjective by presenting a ground-truth-free method for recommending representative images to train FLIM encoders while guiding the user's choice with a visual explanation. Our method suggests new images based on the distance to the previously selected images, using a proposed descriptor that generates different patterns according to filter attention. The user selects the next image among the suggested ones, which repeats until the user's satisfaction. According to our experiments, the proposed method can outperform expert manual selection and an interactive method based on pixel-wise annotation. These results are demonstrated for brain tumor and gastrointestinal parasite segmentation.