BiSe-Unet: A Lightweight Dual-path U-Net with Attention-refined Context for Real-time Medical Image Segmentation

Researchers M Iffat Hossain and Laura Brattain have introduced BiSe-UNet, a novel lightweight architecture designed to solve the critical challenge of deploying real-time medical image segmentation on resource-constrained edge devices. The model specifically targets clinical applications like endoscopy-guided colonoscopy, where detecting polyps in real-time video feeds is essential but computationally demanding. Traditional models often sacrifice either speed or accuracy, creating a deployment bottleneck for embedded medical systems. BiSe-UNet addresses this by proposing a dual-path approach that separates spatial detail preservation from contextual understanding, allowing for efficient processing without compromising diagnostic reliability.

The technical innovation lies in BiSe-UNet's architecture: a shallow spatial path preserves fine-grained details like polyp boundaries, while an attention-refined context path captures broader semantic information. These paths are fused and processed through a decoder built with depthwise separable convolutions, drastically reducing computational load. Evaluated on the public Kvasir-Seg benchmark, the model demonstrates it can maintain accuracy (competitive Dice and IoU scores) while achieving a throughput exceeding 30 frames per second on a Raspberry Pi 5. This performance makes real-time, on-device AI analysis feasible for point-of-care medical diagnostics, potentially transforming how procedures like colonoscopies are performed by providing instant AI-assisted insights without relying on cloud connectivity.