Detection-driven two-stage framework for intraoperative ROSE WSI classification

Background and objective:

Solid pancreatic lesions (SPLs) represent one of the most lethal forms of gastrointestinal malignancies, and Rapid on-site evaluation (ROSE) serves as an important component of intraoperative diagnosis. However, efficient and accurate ROSE slide interpretation remains challenging due to the gigapixel scale of whole-slide images, sparse distribution of diagnostically relevant regions, and the need for real-time feedback.

Methods:

To address challenges, we propose a novel two-stage framework for fast and precise ROSE WSI classification, following the clinical diagnostic workflow of cytopathologists. In the first stage, we design a lightweight Transformer-based object detection network named as RoF DETR, which detects key cell clusters at 5x magnification. To further enhance detection performance, we incorporate domain-specific medical foundation model features and design a multi-scale feature fusion module for effective feature extraction. In the second stage, we design a prototype-guided multiple instance learning network (PG-MIL) based on pseudo-bag augmentation for 20x magnification patch extraction, improving feature discrimination and robustness under class imbalance.

Results:

For comprehensive evaluation, we establish a dedicated ROSE WSI dataset and a cell cluster detection dataset. Our method achieves an $AP@0.5$ of 0.482 in cell cluster detection and an AUC of 92.36% in WSI-level classification. Compared to conventional WSI-level classification pipelines, the proposed framework reduces computational overhead by approximately 100$\times$ and halves the inference time.

Conclusion:

The proposed framework provides a scalable and efficient solution for rapid cytological assessment of ROSE slides, showing potential to support real-time intraoperative decision-making in clinical workflows.