A hybrid CNN–transformer network with cross-level multi-scale fusion and hierarchical attention for brain tumour MRI segmentation

Image segmentation demands a delicate balance between fine-grained spatial precision and global semantic reasoning, yet conventional U-shaped networks struggle to deliver both at once. Three architectural bottlenecks recur across existing designs: convolutional stacks whose effective receptive field expands only sub-linearly with depth; skip connections that indiscriminately forward low-level activations—including background noise—to the decoder; and a latent representation that is built solely from the deepest encoder stage, severing the link between shallow geometric cues and deep semantic abstractions. The proposed network integrates five tightly coupled modules: (1) a Squeeze-and-Excitation–enhanced residual encoder for channel-aware feature extraction; (2) a Transformer bottleneck augmented with 2D sinusoidal positional encoding to model unconstrained long-range dependencies; (3) an Atrous Spatial Pyramid Pooling block sampling the bottleneck at four dilation rates; (4) a Multi-Scale Fusion (MSF) pathway that propagates features from every encoder level into the bottleneck through learnable channel projection and spatial alignment; and (5) an attention-gated decoder in which a Convolutional Block Attention Module performs cascaded channel–spatial refinement at each upsampling step. Training is supervised at three output resolutions and guided by a composite objective combining binary cross-entropy, soft-Dice, focal, and a Laplacian boundary-aware loss. Beyond the architecture itself, we contribute a step-by-step forward-pass algorithm, a parameter count and asymptotic complexity analysis, and a theoretical analysis showing how cross-level fusion broadens the bottleneck’s effective receptive field. Evaluated on a public brain-tumour segmentation benchmark, PRISM-Net is trained end-to-end on real MRI data and assessed through qualitative segmentation results and a post-hoc SHapley Additive exPlanations (SHAP) attribution study; the latter serves as a qualitative sanity check, indicating that the model’s decisions are mildly localised to the lesion region rather than to background structures. A comprehensive quantitative comparison against competing architectures, a component-level ablation, and a cross-validation study form part of an ongoing benchmarking effort and are not claimed in this version.