<p>Deploying computationally intensive deep learning models for medical image segmentation in real-world clinical settings remains challenging, particularly in primary healthcare institutions and resource-limited environments where high-performance servers are often unavailable. To address this issue, we propose APU-net, a lightweight segmentation network that integrates the AT-Lite module and a frequency-domain fusion mechanism to balance computational efficiency and feature representation while maintaining high segmentation accuracy. Extensive experiments on multiple datasets, including publicly available datasets and full-resolution whole-slide images for Spread Through Air Spaces in lung adenocarcinoma, demonstrate that APU-net achieves high segmentation performance and robustness across diverse data sources. Notably, the model enables full-resolution WSI inference within minutes and sustains high processing speed on standard laptop hardware, underscoring its practical utility. Through close collaboration with pathologists, the results were systematically validated via evaluation, case studies, and survival analyses, confirming the model’s clinical reliability. APU-net achieves a strong balance between efficiency and accuracy, making it suitable for real-world deployment on portable devices in clinical workflows.</p>