A3Net: A Lightweight Attention-Guided Encoder–Decoder for Multi-Class Brain Tumor Segmentation with Cross-Dataset Domain Gap Characterization

Accurate multi-class segmentation of brain tumors from T1-weighted contrast-enhanced MRI is essential for surgical planning and treatment monitoring, yet existing high-performance architectures exceed 30 M parameters, limiting deployment in resource-constrained clinical settings. This work introduces A3Net, a lightweight encoder–decoder combining depth-wise separable convolutions, squeeze-and-excitation channel attention, and summation-based skip fusion with a cross-entropy + Dice composite loss. On the cleaned BRISC 2025 benchmark (47 duplicates removed by MD5 audit) under 2-seed × 5-fold cross-validation, A3Net attains 77.79 ± 0.29% mIoU and 87.14 ± 0.17% Dice at 2.11 M parameters, 1.39 GFLOPs, and 8.15 ms latency, achieving statistical parity with Attention U-Net (P=0.334) at 15.0× lower parameter cost. A3Net achieves significant per-class advantages on meningioma (+1.54 pp IoU, P=0.007) and pituitary (+0.42 pp IoU, P=0.016), while Attention U-Net retains a significant advantage on glioma (−1.49 pp IoU, p < 0.001). In-domain evaluation on Figshare yields 74.66 ± 0.50% mIoU (+1.65 pp over Attention U-Net), and a supplementary probe on Akter-Seg confirms efficiency parity with Attention U-Net (83.30 ± 0.44% mIoU vs. 83.21%) at 15.0× fewer parameters across two primary and one supplementary benchmark. Zero-shot cross-dataset transfer exposes a strongly asymmetric domain gap (−12.64 pp BRISC→Figshare; −29.90 pp Figshare→BRISC); multi-source standardized training (BRISC ∪ Figshare, N = 6,655) provides a partial +3.05 pp recovery under target-inclusive training, which is a fundamentally different experimental condition from the target-blind zero-shot setting in which the −29.90 pp gap is measured; the two quantities are therefore not directly commensurable. Class-distribution mismatch, rather than intensity shift, is identified as the dominant residual barrier. Overall, A3Net offers a statistically validated, computationally efficient solution for multi-class brain tumor segmentation with demonstrated cross-dataset generalization.