AI in Medical Imaging: Detection, Segmentation and Device Clearance
AI in medical imaging uses deep learning for tasks like detecting pneumonia on chest X-rays and segmenting structures in biomedical images. The FDA maintains a list of AI-enabled medical devices that have met premarket requirements. Datasets for training require expert annotation and curation.
Detection and Segmentation Tasks
In medical imaging, AI performs two main types of tasks: detection and segmentation. Detection identifies whether a finding, such as pneumonia, is present in an image. Segmentation delineates the exact boundaries of structures or lesions, such as outlining a tumor or an organ. Both tasks rely on deep learning, a machine learning approach that builds complex neural networks with many computational layers, as described by the NIBIB's AI science topic.
A well-known detection example is CheXNet, a 121-layer convolutional neural network trained on over 100,000 frontal-view chest X-ray images to detect 14 diseases, including pneumonia. According to the CheXNet paper on arXiv, the algorithm exceeded the average performance of practicing radiologists on the F1 metric for pneumonia detection.
For segmentation, the U-Net architecture uses a contracting path to capture context and a symmetric expanding path for precise localization. The U-Net paper on arXiv shows that this network can be trained end-to-end from very few images and outperformed prior methods on the ISBI challenge for segmenting neuronal structures in electron microscopic stacks.
Training Data and Annotation
Developing AI for medical imaging requires large, high-quality datasets. The NIBIB's AI science topic explains that datasets must include medical images that are carefully selected, stripped of identifying information, checked for quality, and labeled by clinical experts. This effort is necessary to teach AI tools effectively.
The Medical Imaging and Data Resource Center (MIDRC), supported by NIBIB, provides a large library of X-rays, CT scans, and MRIs from patients across the United States. Researchers use this data to train AI algorithms that could help doctors make earlier diagnoses and more precise outcome predictions, as noted in the NIBIB's AI science topic.
Applications in Cancer Imaging
AI is being applied across cancer research and care, including screening, detection, and diagnosis. The National Cancer Institute's AI and cancer page states that AI can improve the speed, accuracy, and reliability of some cancer screening methods. For example, the FDA has authorized AI-based software that assists pathologists in locating possible cancerous regions in prostate biopsy images.
NCI-supported research has shown that AI imaging algorithms can improve breast cancer detection on mammography and help predict long-term risk of invasive breast cancers. Additionally, NCI scientists have developed a deep learning approach for automated detection of precancerous cervical lesions from digital images, as described on the National Cancer Institute's AI and cancer page.
FDA Clearance and the AI-Enabled Device List
In the United States, AI-enabled medical devices must meet the FDA's premarket requirements before they can be marketed. The FDA's AI-enabled medical devices list is a resource that identifies devices authorized for marketing, providing transparency for healthcare providers and patients. The list includes devices identified primarily based on AI-related terms in their marketing authorization summaries.
The FDA's list is not comprehensive but includes devices that have undergone a focused review of safety and effectiveness, including an evaluation of study appropriateness for the device's intended use and technological characteristics. The FDA plans to explore methods to identify devices incorporating foundation models, such as large language models, in future updates, as stated on the FDA's AI-enabled medical devices list.
Ethical and Governance Considerations
The World Health Organization (WHO) provides guidance on the ethics and governance of AI for health, focusing on large multi-modal models (LMMs). According to the WHO's guidance on large multi-modal models, LMMs can accept one or more types of data input and generate diverse outputs, with predicted wide use in health care, scientific research, and drug development.
The WHO guidance addresses the need for ethical implementation, emphasizing that AI should be used in a scientifically rigorous manner. This is echoed by the National Cancer Institute's AI and cancer page, which notes that AI applications have the potential to advance cancer research and improve health outcomes when implemented ethically.
| Task | Description | Example | Source |
|---|---|---|---|
| Detection | Identifies presence of a finding | CheXNet detecting pneumonia on chest X-rays | CheXNet paper on arXiv |
| Segmentation | Delineates boundaries of structures | U-Net segmenting neuronal structures | U-Net paper on arXiv |
| Classification | Categorizes images or lesions | AI detecting precancerous cervical lesions | NCI AI and cancer page |
| Risk prediction | Predicts long-term outcomes | AI predicting invasive breast cancer risk | NCI AI and cancer page |
Frequently asked questions
What is the difference between detection and segmentation in medical imaging?
Detection identifies whether a finding, such as pneumonia, is present in an image. Segmentation outlines the exact boundaries of structures or lesions, such as a tumor or organ.
How are datasets for medical imaging AI created?
Datasets must be curated, de-identified, assessed for quality, annotated, and labelled by clinical experts, as explained by NIBIB. This effort is necessary to teach AI tools effectively.
What is the FDA's AI-enabled medical device list?
It is a resource that identifies AI-enabled medical devices authorized for marketing in the United States, providing transparency for healthcare providers and patients.
What are large multi-modal models in health?
LMMs are a type of generative AI that can accept one or more types of data input and generate diverse outputs, with predicted wide use in health care, according to WHO.
How is AI used in cancer screening?
AI can improve the speed and accuracy of cancer screening, such as detecting breast cancer on mammograms or precancerous cervical lesions, as noted by NCI.