Quality-Governed Deep Learning Network for Brain Tumor MRI Segmentation

Abstract

The segmentation of brain tumors from magnetic resonance imaging (MRI) is an essential step for computeraided neuro-oncology, treatment planning and follow-up. While U-Net and Attention U-Net achieve effective encoder-decoder representations, the skip connections in these networks can produce low-contrast or noisy responses when the tumour boundary is heterogeneous. In this paper, we introduce a Quality-Governed Deep Learning Network (QGDL-Net) for binary tumour segmentation from FLAIR MRI slices. The proposed model can predict slice quality based on contrast and SNR statistics and fine-tune the encoder features prior to decoding using learnable quality gates. Experiments were conducted using strict patient-wise partitioning, repeated cross-validation with multiple random seeds, and an independent held-out test set to minimize data leakage. QGDL-Net achieved a Dice of 0.9747, IoU of 0.9506, precision of 0.9755, recall of 0.9739, F1 of 0.9747 and accuracy of 0.9982, outperforming U-Net and Attention U-Net. The EigenCAM maps and segmentation overlays provide better tumor localization and interpretable activation around the lesion regions.