Swin-Conv: A Hybrid Swin Transformer and Convolutional Network for Brain Tumor Segmentation

Abstrak

Accurate segmentation of brain Magnetic Resonance Imaging (MRI) plays a crucial role in identifying the boundaries of abnormal regions associated with brain tumors, thereby facilitating more precise diagnosis and treatment planning. Previous studies have proposed automatic segmentation by leveraging deep learning. Convolution-based architectures are effective at spatial localization but are limited in capturing global contextual information, whereas transformer-based architectures can model long-range dependencies but often require substantial computational resources and may struggle to preserve fine-grained spatial details. To address these challenges, this research proposes Swin-Conv, a hybrid U-Net-based architecture consisting of a Swin Transformer encoder to capture the global context of an image and a convolutional decoder to preserve spatial localization during image reconstruction. Furthermore, the effectiveness of standard convolution and Mobile Inverted Bottleneck Convolution (MBConv) employed in the decoder is investigated across four Swin Transformer variants (Tiny, Small, Base, and Large). The experimental results on the public Low-Grade Glioma (LGG) MRI brain segmentation dataset demonstrate that the best performance is obtained by the Swin-Conv model with a standard convolutional decoder and the Swin-S. Comparative experiments with baseline models indicate that Swin-Conv achieves competitive performance with reasonable computational complexity. These findings highlight that Swin-Conv effectively integrates the benefits of a Swin Transformer encoder and a convolutional decoder to generate precise brain image segmentation efficiently, making it suitable for applied medical scenarios.

Identifier

• OpenAlex: https://openalex.org/W7163761063
• DOI: https://doi.org/10.48084/etasr.18002