Learning robust brain tumor segmentation under label corruption and data scarcity

Abstract

Deep learning models for medical image segmentation often struggle with performance issues when datasets are affected by label noise or annotation errors, commonly introduced during manual process or lack of proficiency. These noisy annotations disrupt the loss function, leading to "partially incorrect" gradients that impair the model's learning and overall performance. Additionally, the limited availability of scanned data for training often makes it challenging to develop a robust model. A common approach to address this issue is to leverage similar large annotated datasets. However, differences in dataset distributions can also lead to inconsistencies, introducing erroneous gradients during training and further impacting model performance. To address these challenges, we propose MGR-DAS (Meta-Gradient Reweighting via Direction-Aware Similarity), a novel meta-learning-based approach that can automatically evaluate the reliability of training samples during training using a small, clean subset easily curated from the noisy dataset. Our method quantifies reliability by measuring the cosine similarity between the gradients of noisy training samples and those of the clean subset. Samples with higher gradient alignment are assigned greater weights during training, effectively reducing the impact of noisy labels and improving model robustness. We evaluate our method using three standard metrics for medical image segmentation: the Dice Similarity Coefficient (DSC), the 95th percentile Hausdorff Distance (HD95), and Intersection over Union (IoU). The proposed MGR-DAS achieved an overall 2.4 % improvement in the DSC on the brain tumor segmentation (BraTS, 2021) dataset. Remarkably, even with only 10 clean annotations used in the reweighting algorithm, our method yielded a 28.7 % gain in DSC. In real-world, data-scarce scenarios, our proposed MGR-DAS also improved the overall DSC score by 2.6 % on BraTS pediatric (BraTS-PEDs) and by 1.0 % on BraTS-Africa, demonstrating strong generalizability and robustness. Experimental results confirm that the proposed method reliably identifies noisy data, prioritizes clean data through adaptive weighting, and outperforms existing fine-tuning, curriculum learning techniques, and other meta-learning frameworks commonly employed in classification tasks.