Use Case: Roadmap for Enhanced Diagnostics via AI
Use Case: This use case focuses on the broader implementation of artificial intelligence (AI) in routine diagnostic pathology within a fully digital infrastructure. The overarching goal is to move beyond digitizing slides to actively leveraging AI algorithms for diagnostic purposes. This aims to make pathology diagnostics more objective, faster, and intellectually more satisfying, ultimately benefiting patients by forming the basis for personalized treatment. For breast cancer, specific applications include enhancing the assessment of biomarkers like ER, PR, HER2, and Ki667, and improving the detection of lymph node metastases.
Examples of integrated AI algorithms include:
Ki67 positive nuclei assessment: An AI-based algorithm is included in the Sectra PACS for this purpose.
Mitotic figure recognition: An in-house developed AI algorithm identifies mitoses and mitosis-like objects, allowing pathologists to interactively refine counts.
Workflow stratification: Algorithms like Derm-AI are being evaluated for triaging dermatopathology cases.
Breast cancer package: Algorithms for ER, PR, HER2, Ki67, and lymph node metastases are slated for implementation, designed to run in the background for automatic results.
The development and training of these AI models are facilitated by frameworks like PyTorch and Tensorflow.
Limitations: The implementation of AI in diagnostic practice faces several significant challenges:
Hardware Issues: Running AI algorithms, especially on large whole-slide images (WSIs), requires significant computing power, often necessitating costly local GPU server clusters or external cloud solutions. Cloud solutions introduce security and privacy concerns due to data transfer outside hospital firewalls.
Certification Issues: Algorithms for clinical use must be certified (FDA-approved or IVDR-approved). New European regulations classify AI software as higher risk, requiring assessment by appointed organizations.
Deployment Challenges: Beyond training, deploying models requires robust infrastructure for retraining, data versioning, monitoring, and serving. The rapid evolution of AI technology leads to model drift and technical debt, making it challenging to maintain and port models. There are also ongoing discussions about the trust and explainability of AI models.
Business Case: Digital pathology and AI implementation come with significant costs (acquisition, IT personnel, integration), making a viable business case challenging without external support. While AI may save time, quantifying direct financial savings (e.g., reducing IHC stains) is crucial for justifying investment.
Possible Advancements in Accuracy: The source emphasizes augmented intelligence, where pathologists and AI collaborate to outperform either one alone. For instance, AI-assisted tools can improve the accuracy of HER2 IHC staining interpretation. Future advancements point towards "AI 2.0" or "pathomics", involving the multiscale integration of classical histology, immunohistochemistry (IHC), DNA/RNA sequencing, and spatially-resolved molecular imaging (like spatial transcriptomics and proteomics). This integration will allow AI to make sense of vast and diverse datasets, leading to more comprehensive and accurate diagnoses, particularly for personalized precision medicine.
Future for Digital Pathology: AI is expected to improve the quality of diagnosis, reduce pathologist workload, and potentially lower overall diagnostic costs by reducing the need for extensive IHC staining. The long-term vision is an integrated data infrastructure where advanced machine learning algorithms combine imaging and molecular data to select optimal treatments for each patient, preventing overtreatment and saving costs for society.
Paper: https://www.mdpi.com/2075-4418/12/5/1042?
