ABR-DeepLabV3: Adaptive Boundary-Refined Multi-Backbone DeepLabV3 for MRI Brain Tumor Segmentation++

Accurate brain tumor segmentation from magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, surgical guidance, and quantitative tumor assessment. However, automated segmentation remains challenging due to the heterogeneous tumor appearance, irregular morphology, weak tissue contrast, and ambiguous tumor boundaries. Conventional encoder–decoder segmentation models, including DeepLabV3+, can capture multi-scale semantic information but may still produce boundary inconsistencies when tumor margins are diffuse or visually similar to surrounding tissues. To address this limitation, this study proposes ABR-DeepLabV3+, an Adaptive Boundary-Refined Multi-Backbone DeepLabV3+ framework for MRI brain tumor segmentation. The proposed framework integrates an Adaptive Boundary Refinement (ABR) module after decoder fusion to enhance boundary-sensitive feature representation before final mask prediction. Four encoder backbones, namely InceptionResNetV2, Xception, MobileNetV2, and ResNet50, were evaluated under the same framework using the Figshare Brain Tumor Segmentation (FBTS) dataset consisting of 3064 MRI images from meningioma, glioma, and pituitary tumor classes. The evaluation employed patient-level train–validation–test partitioning together with patient-grouped five-fold cross-validation to reduce subject-level information leakage between data subsets. Experimental results show that ABR-DeepLabV3+ with Xception achieved the best overall test performance, with accuracy of 0.9994, Dice Similarity Coefficient (DSC) of 0.9753, and Jaccard Index (JI) of 0.9519. ResNet50 also achieved competitive performance with DSC of 0.9732 and JI of 0.9478, while MobileNetV2 demonstrated strong class-wise performance for meningioma and pituitary tumors. For the Xception configuration, the component-wise ablation analysis showed that Full ABR achieved higher performance than both plain DeepLabV3+ and an ordinary convolutional refinement control. Relative to plain DeepLabV3+, Full ABR increased DSC from 0.89002 to 0.97530 and JI from 0.80183 to 0.95190, while reducing the 95th-percentile Hausdorff Distance (HD95) from 8.72 to 3.28 pixels and the Average Symmetric Surface Distance (ASSD) from 2.41 to 0.87 pixels. Five-fold cross-validation showed that the Xception-based configuration achieved the highest mean DSC of 0.9017 and mean JI of 0.8258 among the evaluated backbones; however, the lower cross-validation mean relative to the single-split test result indicates sensitivity to patient-level data partitioning rather than unconditional evidence of robustness. Qualitative analysis using segmentation overlays, error heatmaps, and prediction-region heatmaps illustrated representative patterns of contour agreement and localized residual segmentation error without interpreting prediction-derived heatmaps as internal model attention or attribution. Overall, the findings indicate that the proposed ABR design improves within-dataset tumor-region overlap and boundary agreement relative to plain DeepLabV3+ under the evaluated experimental conditions, while external validation and repeated-run statistical evaluation remain necessary to establish broader generalizability and reproducibility.