TL;DR
This paper explores how Artificial Intelligence (AI) can be used to quantify clinical expertise in intracoronary imaging for Percutaneous Coronary Intervention (PCI). The authors propose AI-driven tools that can analyze Optical Coherence Tomography (OCT) images to assess image quality and identify optimal stent landing zones, aiming to assist physicians, especially those with less experience, in making critical decisions during PCI.
What Problem Does This Paper Address?
Percutaneous Coronary Intervention (PCI) is a procedure to open blocked coronary arteries, and intracoronary imaging modalities like Intravascular Ultrasound (IVUS) and Optical Coherence Tomography (OCT) are increasingly recommended to guide these interventions. These imaging techniques provide detailed anatomical information beyond standard angiography, enabling more precise vessel measurements, stent sizing, and disease assessment. However, interpreting these complex images and making critical decisions, such as identifying optimal stent landing zones, requires significant clinical expertise. This expertise is developed over years of practice and involves nuanced judgment of various imaging features, including plaque characteristics, vessel size, and image quality. The challenge is to translate this implicit, expert knowledge into a format that can be readily utilized by less experienced operators, thereby improving the consistency and safety of PCI procedures.
Methods
The paper discusses the development of AI-based software designed to capture and quantify clinical expertise in OCT image analysis for PCI. The core methodology involves training AI models on expert judgments of image quality and stent landing zone suitability.
- Image Quality Assessment: An AI tool has been developed to quantify OCT image quality. Instead of relying on specific technical criteria (like blood clearance or distortion magnitude), this AI is trained on expert assessments of image quality, aiming to mimic the nuanced judgment of experienced physicians. This tool outputs an "index" that reflects the interpretability of the OCT image for clinical decision-making.
- Stent Landing Zone Selection: Another AI tool is presented that quantifies the quality of potential stent landing zones. Similar to the image quality assessment, this AI is trained on expert judgments to identify optimal landing zones, considering factors like vessel size, lumen dimensions, and the presence of plaque. The output is a quantifiable index that helps physicians assess the suitability of different landing zones.
The authors emphasize that these AI tools are built by capturing "expert gestalt" – the collective, often implicit, knowledge and intuition of experienced clinicians.
Datasets
Not explicitly detailed in the paper, but the AI models are trained on deidentified patient data from OCT imaging. The figures showcase examples of OCT runs and their corresponding AI-derived quality indices. The authors mention that the AI systems are trained by expert physicians, implying a dataset of expert interpretations of these OCT images.
Key Insights and Ideas
- Quantifying Implicit Expertise: The most significant insight is the potential for AI to translate the nuanced, often implicit, clinical expertise of experienced interventional cardiologists into quantifiable metrics. This moves beyond simple algorithmic analysis to capturing the "art" of interpretation.
- Democratizing Expertise: By quantifying image quality and landing zone suitability, AI can provide decision support for less experienced operators, potentially reducing variability in PCI outcomes and improving patient safety.
- Beyond Objective Metrics: The AI tools are not based on purely objective, technical criteria but on subjective expert judgment. This is crucial because clinical decisions often involve factors that are difficult to quantify objectively.
- AI as a "Brain" Component: The authors envision AI systems becoming integral parts of imaging software, effectively incorporating the "brains" of expert physicians into the tools used by all operators.
Key Results
- AI-derived Image Quality Index: The AI tool can quantify OCT image quality, providing an index that reflects its interpretability for clinical decision-making, regardless of the type of artifact present. This index can help operators avoid making measurements on poor-quality images or prompt them to repeat OCT runs.
- AI-derived Landing Zone Quality Index: The AI can quantify the quality of stent landing zones, identifying segments with reduced luminal reduction and assessing their suitability for stent placement. This assists in making more informed decisions about stent length and placement.
- Increased Precision and Speed: While not explicitly quantified with numbers in this viewpoint, the authors suggest that these AI tools will lead to increased precision and speed in vessel measurements and decision-making during PCI.
Limitations and Future Work
- Current AI Limitations: The authors acknowledge that current AI software has limitations in the "type of hierarchical intelligence" it offers. While it can answer questions like "where is the vessel?" or "where is the calcium?", it struggles with more complex, context-dependent questions like "can I trust the imaging quality?" or "where exactly should I land the stent in such a diseased calcified vessel?"
- Capturing Nuanced Decisions: The paper highlights that operators instinctively capture interconnected pieces of imaging information and make nuanced decisions based on years of practice. For AI to truly help with increased clinical complexity, it must learn to capture this expert intelligence and output it in usable indices.
- Future Directions:
- Developing AI to capture more complex clinical expertise for PCI decision-making.
- Applying similar AI approaches to IVUS modalities.
- Addressing more relevant clinical questions, such as assessing stent expansion and the need for calcium modification before stent deployment.
- Conducting trials to compare the safety and efficacy of AI-guided PCI versus physician-guided PCI.
- Exploring the broader scope and impact of machine learning on PCI decision-making.
Summary updated