Hannah Yong Wen

Classification performance, computational efficiency, and storage burden (10× versus 20×, TCGA-Glioma)

Despite the impressive performance across a range of applications, current computational pathology (CPath) models face significant diagnostic efficiency challenges due to their reliance on high-magnification whole-slide image analysis. This limitation compromises their clinical utility, especially in time-sensitive diagnostic scenarios and situations requiring efficient data transfer. To address these issues, we developed a computation- and communication-efficient framework called Magnification-Aligned Global-Local Transformer (MAG-GLTrans). The approach substantially reduced computational time, file transfer requirements, and storage overhead by enabling effective analysis using low-magnification inputs. The magnification alignment (MAG) mechanism employed self-supervised learning to bridge the information gap between low- and high-magnification levels by effectively aligning their feature representations. Through extensive evaluation across various fundamental CPath tasks, MAG-GLTrans demonstrated state-of-the-art classification performance while achieving remarkable efficiency gains, including up to a 10.7× reduction in computational time and more than a 20× reduction in file transfer and storage requirements. Furthermore, the versatility of the MAG framework was demonstrated through two significant extensions: (i) its applicability as a feature extractor to enhance the efficiency of any CPath architecture and (ii) its compatibility with existing foundation models, enabling them to process low-magnification inputs with minimal information loss. In a real-world clinical application, computer-assisted telepathology for intraoperative frozen section diagnosis of non–small cell lung carcinoma, MAG-GLTrans effectively recognized distinct tumor morphologies and accurately localized diagnostically relevant regions with low computational and communication costs. Together, these advancements position MAG-GLTrans as a particularly promising solution for time-sensitive applications, especially in the context of intraoperative frozen section diagnosis, in which both accuracy and efficiency are paramount.Significance:MAG-GLTrans is a general-purpose whole-slide image analysis architecture designed for computational pathology applications, including cancer subtyping, metastasis detection, and mutation prediction, that simultaneously optimizes diagnostic capability, computational efficiency, and communication burden.

Visualization of the features extracted by MAG (CAMELYON16)

Comparison on different field of views (FOV)

Comparison with an efficiency baseline ZoomMIL

Classification performance comparing with pathology foundation models (5× versus 20×)

Classification performance, computational efficiency, and storage burden (5× versus 20×, TCGA-Glioma)

Classification performance, computational efficiency, and storage burden (10× versus 20×, TCGA-NSCLC, TCGA-Glioma, CAMELYON16)

Supplementary Figures S1-S5 Supplementary Tables S1-S5