EvoBTSeg: Evolutionary Architecture Search for Compact and Precise Heterogeneous Brain Tumor Segmentation

Accurate and efficient segmentation of heterogeneous brain tumors requires neural architectures that can simultaneously capture diverse tumor subregions with varying appearance, scale, and boundary characteristics while remaining computationally compact. However, manually designed networks with fixed architectures often struggle to balance expressive feature representation and model efficiency across such heterogeneity. To address this challenge, we propose EvoBTSeg, an evolutionary architecture search framework that automatically discovers compact and effective network configurations designed for heterogeneous brain tumor segmentation. A customized U-Net supernet is defined with parameterized encoder, bottleneck, and decoder components, enabling flexible control over network depth, filter size, activation function, and dropout rate. The encoder extracts hierarchical representations through convolution, ReLU activation, max-pooling, and dropout to accommodate heterogeneous tumor patterns, while the bottleneck refines global contextual features via convolution and weighted feature aggregation. The decoder progressively restores spatial resolution using transposed convolution and feature concatenation, preserving fine-grained boundary information critical for accurate delineation of small and diffuse tumor subregions. An evolutionary search strategy based on crossover and mutation is employed to explore the architectural search space in a data-driven manner, selecting architectures that balance contextual modeling and boundary precision. Experimental results on the BraTS 2021 dataset demonstrate that EvoBTSeg identifies a compact architecture with only 0.12 M parameters, achieving a dice similarity coefficient (DSC) of 92.83% and a Hausdorff distance (HD95) of 1.38 mm. Furthermore, EvoBTSeg demonstrates reliable generalization, maintaining competitive segmentation performance on BraTS 2020 (DSC 89.37%, HD95 1.33 mm) and the ACDC dataset (DSC 88.03%, HD95 1.41 mm). These results indicate that EvoBTSeg effectively discovers lightweight yet accurate architectures for heterogeneous brain tumor segmentation, highlighting its potential for practical clinical deployment.