Digital pathology and AI implementation in pathology practice

This thesis documents the comprehensive implementation of digital pathology and artificial intelligence in clinical diagnostics, tracing the evolution from traditional microscopy to fully digital workflows enhanced by AI-powered diagnostic tools. The digital transformation began with establishing a digital pathology system at UMC Utrecht, focusing on the technical infrastructure required for whole-slide imaging, including scanner integration, data storage solutions, and laboratory information system connectivity. The implementation prioritized open standards to ensure seamless communication between hospital and laboratory systems. A pilot study examining digital slides for mitosis counting in breast cancer revealed reasonable reproducibility, though with slight underestimation in high-proliferation tumors. The transition to fully digital primary diagnostics was completed within six months in 2015, making UMC Utrecht among the first laboratories to achieve a comprehensive digital archive. Most pathologists and residents expressed high confidence in the system, though certain specialties like pediatric, hemato-, and neuropathology showed slower adoption rates. The digital workflow presented challenges in distinguishing microorganisms and mitoses in scanned slides, but offered improved ergonomics compared to traditional microscopy. Notably, the implementation reduced mean turnaround times by nearly 7%, saving approximately half a day per case. The COVID-19 pandemic highlighted the strategic value of digital pathology infrastructure. Remote work capabilities enabled pathologists, residents, and laboratory staff to maintain effective workflows while adhering to social distancing requirements. Survey results indicated that most pathologists found remote work as effective as on-site operations, with reduced stress levels and increased flexibility. The laboratory achieved 90-95% digital sign-out rates, occasionally reverting to glass slides for cases requiring higher diagnostic confidence. Significant milestones included implementing the first routine image analysis tool for Ki67 scoring and establishing integration with a national platform for whole-slide image exchange between Dutch pathology laboratories. The artificial intelligence implementation focused on tumor proliferation assessment through organized challenges and clinical validation studies. The TUPAC16 challenge addressed automated tumor proliferation scoring from whole-slide images, achieving promising results with quadratic weighted Cohen's kappa scores of 0.567 for mitotic scores. The MIDOG 2021 challenge specifically targeted scanner-agnostic mitotic figure detection, with the winning algorithm achieving an F1 score of 0.748, outperforming human experts. Clinical validation studies demonstrated the practical application of AI in pathology practice. A convolutional neural network for pigmented skin lesion classification achieved 98% accuracy for melanoma detection and 92% for nevi classification. Breast cancer grading validation showed strong correlations between traditional light microscopy, digital imaging, and AI-assisted counting methods, with correlation coefficients ranging from 0.77 to 0.95. The culminating validation study on 912 breast cancer patients with long-term follow-up data confirmed that AI-based mitotic counting retained prognostic accuracy equivalent to traditional methods. The AI system successfully integrated into clinical PACS systems, demonstrating the feasibility of implementing validated AI algorithms in routine diagnostic workflows. This comprehensive approach illustrates how digital pathology infrastructure serves as the foundation for AI integration, ultimately enhancing diagnostic accuracy, improving workflow efficiency, and maintaining prognostic reliability while addressing the practical challenges of clinical implementation.