Evaluation of boundary-aware UNet++ for low-grade glioma segmentation

Abstract

Automated brain tumour segmentation remains challenging because of diffuse boundaries, class imbalance, and variation in lesion appearance. We evaluated a boundary-aware low-grade glioma segmentation framework using UNet++ with an ImageNet-pretrained EfficientNet-B4 encoder and a composite objective including an SDF-guided confidence regulariser. Evaluation used a held-out 22-patient internal test cohort, controlled component and architecture ablations, a three-seed SDF comparison, and zero-shot testing on three external BraTS cohorts. Internally, mean tumour-positive slice DSC was 0.8320, whole-cohort DSC was 0.9086, and mean patient-level DSC was 0.8731; mean slice HD95 was 7.66 pixels. The SDF regulariser showed favourable average multi-seed effects. Under the primary replicated-FLAIR external protocol, mean patient-level DSC was 0.6194 for TCGA-LGG, 0.6007 for full BraTS 2021, and 0.3862 for BraTS-Africa. These findings support encouraging internal performance but limited cross-dataset robustness.