Answers are grounded in this paper and cite their sources.
Accurate brain tumour segmentation is fundamental for neurosurgical planning, radiotherapy guidance, and prognostic assessment. Existing models depend on multimodal 3D MRI acquisitions and computationally intensive architectures that limit clinical deployment. The survival prediction models depend on complex radiomic or multi-omic features with limited interpretability. The contribution of the study is a FLAIR-only 2.5D contextual segmentation design that injects limited through-plane information into an ImageNet-pretrained ResNet50 encoder and couples it with an attention-modulated lightweight UNet + + decoder, to improve volumetric context modelling without requiring full 3D convolution. The proposed segmentation model uses a five-slice 2.5D FLAIR neighbourhood processed by an Adaptive Inter-Slice Context Fusion module that learns content-dependent slice relevance before ImageNet-compatible projection and ResNet50 encoding. The resulting multiscale features are reconstructed using an attention-modulated UNet + + Lite decoder. A hybrid loss function combining categorical cross-entropy and focal Tversky loss is used to mitigate severe class imbalance, while test-time augmentation and conditional largest-connected-component filtering improve robustness and prediction consistency. The AICF-enabled model achieved a whole-tumour Dice score of 0.899, IoU of 0.817, sensitivity of 0.910, and Surface Dice at 3 mm of 0.968 with physical-space HD95 of 0.96 mm. Voxel-wise PR-AUC and ROC-AUC were 0.895 and 0.9988, respectively. Temperature-scaled calibration and uncertainty analyses showed improved confidence–outcome agreement and higher uncertainty for segmentation errors. For prognostic evaluation, age and verified resection status were used to construct the primary interpretable clinical survival baseline through Kaplan–Meier estimation and penalized Cox proportional-hazards modelling. A secondary exploratory analysis examined whether selected segmentation-derived tumour burden and morphology descriptors show incremental prognostic information. Apriori association-rule mining uncovered clinically intuitive age-based risk strata with clearly separated survival trajectories. The findings indicate that a FLAIR-only AICF 2.5D ResNet50–UNet + + Lite architecture delivers accurate, boundary-aware, and resource-efficient brain tumour segmentation. The simple clinical variables combined with explainable rule mining shows transparent survival stratification. This study depends on an internal BraTS 2019 cohort without independent external validation. The findings interpreted as benchmark-level evidence and not as evidence of clinical deployment readiness.
No takes yet. Share an insight, caveat, or question.
Wankhede et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: