Consensus-based iterative meta-pseudo-labeling for deep semi-supervised learning

A known issue that hinders the development of deep learning models is the need for accurate annotation of a large quantity of samples – a time-consuming, labor-intensive, and error-prone task. This limitation is particularly critical in areas where data annotation requires expert knowledge. Semi-supervised learning methods, such as pseudo-labeling, can alleviate the problem by capitalizing on both limited labeled and plentiful unlabeled data; nonetheless, state-of-the-art methods often require pre-trained encoders and validation sets to deliver effective solutions. Herein, we introduce a teacher-student-based iterative meta-pseudo-labeling approach, named consensus Deep Feature Annotation (cons-DeepFA), that enables the training of custom Convolutional Neural Networks (CNNs) from small quantities of labeled samples without reliance on pre-trained encoders and validation sets. cons-DeepFA explores Feature Learning from Image Markers (FLIM) to initialize the filters of a target CNN (student) from minimal data annotation – i.e., user-drawn markers on discriminative regions of a few selected images per class. During each of a few iterations, the latent space of the student's last dense layer is non-linearly projected onto a two-dimensional space for downstream label propagation via an optimum-connectivity-based approach (teacher); afterward, the student is re-trained using pseudo-labeled samples selected by the proposed consensus mechanism, which jointly improves the latent space, its projection, and the student's generalization ability as iterations progress. This strategy was recently introduced with pre-trained encoders by selecting the most confident pseudo-labeled samples to re-train the student. While building on previous methods, cons-DeepFA presents two key contributions. It (i) incorporates FLIM to enable training a custom CNN from scratch with faster convergence, improving its generalization ability, and (ii) introduces a consensus-based procedure over multiple iterations that selects more accurately pseudo-labeled samples for re-training the CNN. Lastly, cons-DeepFA is evaluated on five challenging biological image datasets, demonstrating its effectiveness and competitiveness when compared to seven state-of-the-art methods from four semi-supervised learning paradigms.