Abstract
Abstract
Brain tumor segmentation from multimodal magnetic resonance imaging (MRI) is a central component of treatment planning and disease monitoring, but reliable voxel-wise annotation is costly and difficult to obtain. This work presents a pilot evaluation of UASC-Net Lite v2, an uncertainty-aware extension of a compact 3D U-Net baseline for binary tumor segmentation from multimodal BraTS 2021 MRI patches. The model combines a 3D encoder–decoder segmentation pathway with attention-, prototype-, and uncertainty-related auxiliary outputs, enabling spatial confidence inspection alongside mask prediction. Experiments were conducted on a reproducible 16-case pilot subset (13 training and 3 validation cases), using 64×64×64 patches and eight training epochs. After validation-set threshold calibration, the proposed model achieved a pooled Dice score of 0.7381 and IoU of 0.5849, compared with 0.7297 and 0.5744 for the TinyUNet3D baseline. UASC-Net Lite v2 improved sensitivity from 0.7681 to 0.8609, while specificity and precision changed from 0.9685 and 0.6949 to 0.9532 and 0.6460, respectively. Qualitative visualizations show the segmentation outputs together with uncertainty and attention maps. These results should be interpreted as a pilot study rather than a full benchmark: threshold calibration was performed on the validation patches and the cohort is small. Nevertheless, the experiment provides reproducible evidence that uncertainty-aware modeling can improve lesion recall and overlap under limited-data conditions.