Autonomous labeling of surgical resection margins using a foundation model

Area of Science:

• Digital Pathology
• Computational Pathology
• Surgical Pathology

Background:

• Accurate assessment of resection margins is critical for patient outcomes in surgical pathology.
• Current physical inking methods for margin delineation are variable and can be obscured by artifacts like cautery.
• Standardization and automation are needed for reliable margin assessment in digital pathology workflows.

Purpose of the Study:

• To develop and validate a virtual inking network (VIN) for autonomous localization of surgical cut surfaces on whole-slide images.
• To reduce dependence on physical inks and standardize the review of surgical margins.
• To provide a reproducible, ink-free method for margin delineation in digital pathology.

Main Methods:

• A virtual inking network (VIN) was developed using a frozen foundation model for feature extraction and a two-layer multilayer perceptron for patch-level classification.
• The model was trained on ~2 TB of hematoxylin and eosin (H&E) stained human tonsil tissue slides (~120 slides) with pathologist-provided annotations.
• Blind testing was performed on 20 previously unseen slides to evaluate VIN's performance in margin delineation.

Main Results:

• The VIN successfully produced coherent margin overlays that qualitatively aligned with expert annotations across serial sections.
• Quantitative analysis showed a region-level accuracy of approximately 73.3% on the test set.
• Errors were localized and did not disrupt the overall continuity of the whole-slide margin map, indicating VIN captures histomorphology related to cautery.

Conclusions:

• The virtual inking network (VIN) demonstrates the capability to autonomously delineate surgical margins from whole-slide images.
• VIN offers a reproducible and ink-free approach to margin assessment, suitable for integration into digital pathology.
• This technology has the potential to enhance the standardization and efficiency of margin evaluation in routine pathology practice.