Dual examiner consistency learning with dynamic receptive fields and class-balance refinement for Barely-supervised brain tumor segmentation

Abstract

Brain tumor segmentation from magnetic resonance imaging data is a critical task in medical image analysis, yet it remains challenging due to the complex and heterogeneous nature of tumors, as well as the scarcity of labeled data. In this study, we present a novel barely-supervised learning (BSL) framework for accurate brain tumor segmentation, specifically designed to address the limitations imposed by limited labeled data. Our approach introduces two key components: the dual examiner strategy (DES) and the dynamic receptive convolutional network (DRCN). The DES combines consistency learning with adversarial training to make efficient use of both labeled and unlabeled data. This strategy encourages the model to learn robust and generalized features from unlabeled data while simultaneously ensuring high accuracy through labeled data. To further enhance segmentation performance, the DRCN module adaptively adjusts the receptive fields during feature extraction, enabling the model to better capture tumor boundaries, which are often irregular and spatially varied. Additionally, we propose a novel class-balancing refinement (CBR) loss to address the problem of class imbalance commonly encountered in tumor segmentation tasks. This loss function dynamically reweights the classes during training, allowing the model to focus on underrepresented regions, thereby improving segmentation accuracy for smaller tumor areas. We validate our approach on the BraTS 2019, 2020, and 2021 datasets, achieving significant improvements in segmentation performance with minimal labeled data. Our results demonstrate that the proposed method outperforms existing techniques in terms of both accuracy and robustness, offering a promising solution for brain tumor segmentation in data-scarce scenarios.