Paige BLN AI cuts pathologist reading time 55% and raises breast cancer lymph node metastasis sensitivity from 81% to 93%

Detection of breast cancer lymph node metastases is a tedious, time-consuming task where pathologist sensitivity is suboptimal, particularly for small deposits such as isolated tumor cells (ITCs). Research shows that when subspecialized breast pathologists re-review lymph nodes originally diagnosed by non-specialists, 24% of patients are assigned a different—usually higher—nodal stage, highlighting a meaningful diagnostic accuracy gap with direct staging implications.

Paige BLN, a deep learning tumor detection system trained on over 32,000 whole slide images (WSIs) from more than 8,000 patients using weakly supervised multiple instance learning, was integrated into a digital pathology viewer. The system provides a binary cancer/no-cancer classification and highlights the highest-probability suspicious region on positive slides, with pathologists able to toggle the AI overlay on demand. Three board-certified pathologists at New England Pathology Associates evaluated 167 breast sentinel lymph node WSIs—enriched for challenging cases including ITCs and small micrometastases—in a randomized crossover design with and without AI assistance.

Pathologists using Paige BLN reduced average reading time from 129 seconds to 58 seconds per slide, a 55% efficiency gain (P<0.001) that applied equally to benign and malignant slides. Overall group sensitivity improved from 81.2% to 93.2%; two of three individual pathologists achieved statistically significant sensitivity gains (72.5%→94.2% and 78.3%→92.8%, both P≤0.006). ITC sensitivity nearly doubled from 46.2% to 79.5%. Specificity was maintained at approximately 97% in both reading modes, confirming no increase in false positives.

• AI assistance delivers the largest accuracy benefit for the hardest-to-detect lesion subtypes (ITCs and small micrometastases), where unaided human sensitivity is lowest and staging consequences are highest.
• The 55% reduction in reading time held across both benign and malignant slides, demonstrating workflow efficiency gains regardless of case outcome.
• Specificity was preserved with AI assistance, meaning faster reads did not trade accuracy for speed.