AI-enhanced precision: boosting pathologist confidence in HER2 ultra-low diagnosis through intelligent re-screening

Accurate assessment of Human Epidermal Growth Factor Receptor 2 (HER2) immunohistochemistry (IHC) expression, particularly the precise identification of "HER2 ultra-low" status, is critical for guiding antibody-drug conjugate (ADC) therapies, such as trastuzumab deruxtecan (T-DXd) in breast cancer. Given the difficulty and poor consistency in interpreting HER2 ultra-low expression, this study evaluated an artificial intelligence (AI) tool for assisting pathologists in diagnosing HER2 ultra-low expression in breast cancer. The analysis included 188 breast cancer cases (376 whole-slide images), consisting of 136 matched core needle biopsy (CNB)-surgical excision (SE) pairs from patients without neoadjuvant therapy (NAT) and a separate cohort of 52 matched CNB-post-NAT SE pairs with HER2 IHC scores of 0 or 1+. All slides were scanned using the Roche VENTANA DP 600 scanner. The resulting whole-slide images (WSIs) were subsequently analyzed and scored for HER2 IHC expression using AI assistance via the Roche Navify^® Digital Pathology platform. Expert consensus diagnosis served as the ground truth. Nine pathologists of varying experience re-evaluated their assigned slides with and without AI assistance, assessing changes in diagnostic accuracy, efficiency, and confidence. Additionally, an external cohort of 89 slides from two independent hospitals was used to validate the generalizability of the AI-assisted HER2 interpretation. Among 92 HER2 ultra-low slides (40 biopsy and 52 SE sections) identified by experts, AI assistance significantly improved pathologists' detection performance, increasing the F1-score from 0.78 to 0.98, precision from 0.39 to 0.91, and recall from 0.72 to 0.96 against expert consensus. In non-NAT matched pairs, AI improved the concordance between CNB interpretation and SE ground truth from 0.37 to 0.48, primarily by reducing HER2 underestimation. In post-NAT SE specimens, AI increased classification consistency from 0.63 to 0.92. Furthermore, AI assistance significantly elevated diagnostic agreement among pathologists, reduced interpretation time, and increased the proportion of high-confidence assessments. These findings demonstrate that AI support enhances the accuracy, consistency, and efficiency of HER2 ultra-low identification across biopsy, resection, and post-treatment specimens, potentially improving patient selection for novel ADC therapies. With AI support, pathologists can deliver more precise and confident diagnoses of HER2 ultra-low breast cancer.