AI has been a popular term in radiology for many years, with most of the field’s AI applications achieving U.S. Food and Drug Administration (FDA) approval before the widespread use of GPT-3 and GPT-4 in 2023. For many, machine learning's potential is more than just talk; the power of well-trained, image-augmented AI has already been demonstrated as a useful tool for quality control, disease tracking, tumor segmentation, and making prognosis predictions.
AI will continue to be a driving force in 2024, leading to gradual improvements in effectiveness of care and resulting in profound impacts on the patient journey. Read on for our views on how standardized, adaptable infrastructure for medical imaging AI will ultimately improve the efficiency and precision of radiological assessments.
1. Augmentative AI, Not Autonomous
AI has been a popular term in radiology for many years, with most of the field’s AI applications achieving U.S. Food and Drug Administration (FDA) approval before the widespread use of GPT-3 and GPT-4 in 2023. For many, machine learning's potential is more than just talk; the power of well-trained, image-augmented AI has already been demonstrated as a useful tool for quality control, disease tracking, tumor segmentation, and making prognosis predictions.
AI will continue to be a driving force in 2024, leading to gradual improvements in effectiveness of care and resulting in profound impacts on the patient journey. Read on for our views on how standardized, adaptable infrastructure for medical imaging AI will ultimately improve the efficiency and precision of radiological assessments.
1. Augmentative AI, Not Autonomous
The narrative of radiology AI has evolved from the notion of autonomous systems replacing expert radiologists, to an assistive approach. AI can take over administrative tasks that radiologists don’t need to perform and augment the rest. Gaps in medical records and communication remain an all-too-common safety risk for patients and are easy challenges for large language models to solve, such as summarizing medical histories for informed assessments, communicating incidental findings to patients, and scheduling follow-ups.
AI can be leveraged to make incremental improvements at every stage of the medical imaging pipeline, beyond the tasks well-suited for a large language model. Imaging AI models can inform contrast or radiation dosimetry, identify crucial information about patients such as age or underlying conditions for emergency response triaging, run scan quality assessments to automate transformations, flag equipment or patient positioning concerns, and identify lesions in a scan for an expert eye to confirm.
These microaugmentations at every step of the imaging processing journey may mean seconds saved for an individual undergoing treatment. Collectively, these microimprovements reduce years of manual data processing work across an integrated health system, allowing many more patients to receive radiology results faster. This shift will not only enhance accuracy at every step of the imaging pipeline but will also reduce the overall workload on radiologists and allied staff.
2. Responsible AI in Healthcare
The radiology field is coming together to define best practices for the development of trustworthy AI, with RSNA, the American College of Radiology (ACR), the European Society of Radiology (ESR), and two other professional societies recently issuing a joint statement with guidelines for the transparent development and introduction of AI into clinical practice.
Transparency measures include disclosing study design, sample size, demographic characteristics, equipment specifications, imaging modality, target population, and target use case. Consistently reporting the original intent of algorithms and model engineering parameters will only bolster trust in niche algorithms and highlight areas that are lacking specialized models.
While it can be burdensome to introduce systems to verify, track, and explain a model’s development lifecycle and ensure that the model’s adoption poses significant benefits as compared to existing tools, this effort is not only crucial for widespread adoption but also for responsible reliance on AI in practice. Overreliance on AI can lead to poorer performance than radiologists acting without AI tools or even autonomous AI alone. Training, uncertainty visualizations, cognitive checks, and enterprise AI infrastructure that surfaces model performance scores and the characteristics of source data can be effective to counter overreliance or underreliance on AI.
3. Multimodal, Patient-Centric Registries Accelerate Clinical Trial Enrollment Decisions
Once efficiency gains are realized by streamlining noninterpretive, “mundane” imaging curation processes with AI, health systems can begin to build “gold standard” stores of radiology data that are pre-labeled and tagged. These living repositories include years of imaging and representative samples of thousands of patients with multiple backgrounds and indications, who may be eligible for downstream enrollment in life-saving clinical trials.
While administrations can realize time savings or reduced risk from introducing AI algorithms into care, the most important asset to revolutionizing care is having organized, multimodal imaging data. When clinicians have searchable repositories at their fingertips, a health system’s owned data is evergreen for connecting patients to clinical trials or for model building, anywhere and at any time. Singular pieces of a patient’s medical record, such as a radiology report or a scan from a single time point, are often incomplete or miss significant portions of a patient’s timeline.
Robust patient registries that include text-based data and medical imaging are normally very time-intensive to build, requiring manual extract, transform, and load (ETL) methods to combine data over weeks or months. By leveraging solutions that automatically combine this data into visual dashboards and support natural language queries, providers have a complete view of a patient’s medical history, which is essential for informed treatment plans and a precursor to external collaborations with real-world evidence.
Pharmaceutical organizations are already testing the promise of AI to optimize clinical trial inclusion and exclusion criteria and prevent costly failures. By analyzing large, multimodal, real-world data sets, inclusion criteria can be refined to only the subjects most likely to respond to treatment. The future of clinical trial enrollment decisions, and of large foundational medical models, depends on correlating medical imaging data sets with EMR data and aggregating these data points into cohorts that enhance the ability of AI to support clinical correlation.
Fortunately, solutions exist today that can tie together de-identified medical imaging with associated tabular data. Eventually, clinicians will be able to identify phenotypes via deep phenotyping, turning images from associated data on the medical record into quantitatively useful tools.
Unlike genotype, which prescribes the way organ systems and body structures will react to a life event, medical imaging phenotypes represent the cumulative result of a patient’s life history. Medical imaging accumulates that pixel-level evidence more comprehensively than any other format. Deep phenotyping will enable the prognostication of disease measures in ways not previously thought possible.
4. Healthcare AI Demands Cloud-based Enterprise Imaging
Organizations are increasingly embracing cloud enterprise imaging as a strategic move to prepare for the implementation and development of AI applications. This shift allows the security, scalability, and accessibility of medical imaging data necessary for developing accurate and trusted models.
About Flywheel
Flywheel is a global medical imaging data management and AI development platform that empowers clinicians to identify cohorts, curate data, and train models to nurture trusted AI. We envision a world where each insight from medical imaging is precise and impactful, and the potential of imaging data is harnessed for AI.