RSAA-UNet: A Residual Squeeze-and-Excitation Attention U-Net with Atrous Context and Focal-Tversky Supervision for Lower-Grade Glioma Segmentation on FLAIR MRI under a Patient-Wise Evaluation Protocol

Abstract: Outlining lower-grade glioma (LGG) on FLAIR MRI is a prerequisite for volumetric follow-up and radiogenomic work, yet many published U-Net pipelines for the public TCGA-LGG collection do not state, or do not enforce, a clean separation of patients between training and testing. This paper describes RSAA-UNet, a 2-D encoder- decoder that combines residual double-convolution blocks, squeeze-and-excitation (SE) recalibration, additive attention gates on all four skip connections, an atrous spatial pyramid pooling (ASPP) module at the bottleneck, and deep supervision at three decoder levels, trained with a Focal-Tversky plus binary cross-entropy objective. Ten networks (three baselines, two RSAA-UNet variants and five single-component ablations) were trained under one protocol on a patient-wise 78/16/16 split of 110 patients (3,929 slices) and evaluated on 16 unseen patients (498 slices, 182 containing tumour). Flip-based test-time augmentation and small-component removal were examined as inference options. On tumour-containing slices the full RSAA-UNet with both inference steps reached a Dice coefficient of 0.715 (patient-bootstrap 95% CI 0.651–0.771), against 0.714 for U-Net and 0.743 for the strongest baseline, a residual U-Net; no difference was significant at the patient level after Holm correction. The proposed objective lowered the false- positive rate on tumour-free slices to 2.5%, compared with 4.4–6.0% for the baselines, and the ablation pointed to residual connections and SE recalibration as the most useful components, while attention gates and ASPP gave no measurable benefit here. Training of the full model also diverged after epoch 20, so its weights come from an early checkpoint. The Discussion explains why a strict patient-wise protocol gives lower Dice values than those usually quoted for this dataset, compares the results with five published studies, and lists the changes needed before the architecture can be judged conclusively.

Keywords: lower-grade glioma; brain MRI segmentation; U-Net; attention gate; squeeze-and-excitation; atrous spatial pyramid pooling; Focal-Tversky loss; patient-wise evaluation