A Comprehensive AI-Based Approach in Classifying Breast Lesions: Focusing on Improving Pathologists' Accuracy and Efficiency

Abstract

Background: Accurate classification of breast lesions is essential for effective clinical decision-making and patient management. In this study, we evaluated an artificial intelligence (AI) solution to classify whole slide images (WSIs) of breast lesions.

Methods: We analyzed a cohort of 104 breast cases, including 20 invasive carcinomas, 4 microinvasive carcinomas, 15 ductal Carcinoma in situ (DCIS) cases, and 65 lobular neoplasia/benign cases. The AI's performance was compared with the ground truth established by breast pathologists.

Results: For invasive carcinoma, it achieved an area under the curve (AUC) of 0.976, with sensitivity and specificity of 91.7% (84.4%, 95.4%) and 95% (88.0% 97.3%) respectively. For DCIS, the AUC was 0.976, with sensitivity and specificity of 93.3% and 96.6%. For lobular neoplasm, it achieved an AUC of 0.953, sensitivity of 94.1%, and specificity of 95.8%. The AI also performed well in detecting microcalcifications, with an AUC of 0.925 and sensitivity of 95%. Pathologists' diagnostic accuracy improved from 97.1% to 100% with AI support (303 vs. 312 accurate case reads per arm). Additionally, the AI use significantly enhanced the pathologists' efficiency, reducing their review time by an average of 16.5% across the 3 pathologists and led to a 33% reduction in immunohistochemistry usage.

Conclusion: This study highlights the potential of AI in breast lesion classification, demonstrating high sensitivity, specificity, and efficiency, and supports its integration into routine pathology practice.

Keywords: Artificial intelligence; Ductal carcinoma in situ; Invasive breast carcinoma; Lobular neoplasia; Microcalcification.