Lentis/AI & Medical Imaging

AI and Medical Imaging

Introduction

Artificial Intelligence (AI) is revolutionizing medical imaging by employing machine learning algorithms and advanced analytics to diagnose, predict, and manage medical conditions with unprecedented precision. By automating repetitive tasks, enhancing diagnostic accuracy, and uncovering subtle patterns undetectable to clinicians, AI has become a vital tool in imaging techniques like X-rays, MRIs, CT scans, and ultrasounds. Its impact spans multiple specialties, including radiology, pathology, ophthalmology, cardiology, and dermatology.

AI’s integration into clinical workflows improves efficiency and accuracy and facilitates earlier disease detection, personalized treatments, and increased accessibility to healthcare, particularly in underserved regions. From identifying early-stage tumors to guiding skin cancer assessments, AI is transforming the diagnostic landscape. However, its rapid adoption also introduces complex regulatory, ethical, and trust challenges that demand careful consideration.

With its far-reaching applications and potential to enhance patient outcomes, AI in medical imaging represents a pivotal advancement toward more equitable, efficient, and effective healthcare systems.

History

The history of artificial intelligence in medical imaging reflects its transformation from early rule-based systems to advanced deep learning applications that revolutionize healthcare. In the 1970s, systems like INTERNIST-1 and MYCIN used algorithms to assist in diagnosing complex conditions, with MYCIN specifically recommending antibiotics based on backward-chaining logic. By 1986, DXplain advanced diagnostic support by generating differential diagnoses from symptom inputs, initially covering 500 conditions and later expanding to thousands, demonstrating AI’s growing impact on clinical workflows.

The completion of the Human Genome Project in 2003 provided a wealth of genetic data, enabling AI to analyze genetic markers and integrate into precision medicine. This laid the foundation for tailoring treatments to individual patients. A significant leap occurred in 2012 with the rise of deep learning technologies like AlexNet, which revolutionized image analysis using convolutional neural networks. This enabled precise detection of anomalies in medical imaging, setting the stage for widespread AI adoption in radiology and diagnostics.

In 2017, the FDA approved the first AI-powered diagnostic device, marking a major regulatory milestone and highlighting AI’s reliability in clinical practice. Around this time, AI was integrated into surgical robots and real-time imaging systems, enhancing decision-making and supporting minimally invasive procedures. The field advanced further in 2020 when Google DeepMind’s AlphaFold accurately predicted protein structures, addressing a fundamental challenge in biology and expanding AI’s applications into biomedical research and drug discovery.

These milestones showcase AI’s evolution from assisting diagnostics to enabling predictive and research capabilities, fundamentally transforming healthcare with each innovation.

Applications

AI has imaging applications across a variety of medical specialties. For example, in the field of radiology, the company Koios has an ultrasound (US) imaging software specialized for the breast and thyroid to assist in identifying malignancies, along with offering diagnostic suggestions to help generate more accurate diagnoses (Densford, 2019). They aim to reduce unnecessary biopsies and expedite treatment decisions through the standardization and improved accuracy that the AI software can provide.

Related to the field of cardiology, Arterys is a company that provides AI-powered tools for cardiac imaging (Fukushima, 2023). The program enables precise quantification of blood flow and heart function from MRI scans. It also has a cloud-based platform that facilitates real-time analysis and collaboration among healthcare professionals, enhancing radiologist workflow while ensuring interoperability.

Paige.AI is a tissue sample analysis tool used in the field of pathology (Paige, 2024). It utilizes AI to assist pathologists in analyzing digital slides, improving the detection of cancerous tissues and streamlining workflow.

IDx-DR is an FDA-approved AI system that is commonly used in the field of ophthalmology to analyze retinal images and detect diabetic retinopathy (Savoy, 2020).  The company strives to increase accessibility to diabetic retinopathy screening, facilitating early detection and treatment to prevent vision loss (FDA, 2018).

SkinVision is a mobile application that uses AI to assess skin lesions for signs of melanoma and other skin cancers, providing users with risk assessments and recommendations (SkinVision, 2024). Their mission is to empower individuals with a tool for early skin cancer detection, promoting timely medical consultation and treatment.

All of these applications share common advertisements like a faster diagnosis, improved accuracy, earlier detections, improved efficiency, and potential cost saving and better patient outcomes.

Key Technologies

AI-driven medical imaging enhances diagnostics through advanced technologies, driving adoption and transformative potential while tackling privacy, scalability, and reliability challenges.

Convolutional Neural Networks (CNNs)： CNNs are the foundation of AI in medical imaging, excelling in tasks like image classification, segmentation, and anomaly detection. These networks analyze pixel-level details in images, enabling precise identification of patterns, such as early-stage tumors or subtle structural abnormalities.

Data Augmentation and Preprocessing： To ensure robust model performance, data augmentation artificially expands training datasets by introducing variations, such as rotations or noise. Preprocessing steps, including normalization and noise reduction, improve image clarity, ensuring high-quality inputs for AI models.

Federated Learning： Federated learning allows decentralized AI training across multiple institutions while maintaining patient data privacy. This technology enables models to learn from diverse datasets without directly sharing sensitive information, improving generalizability across populations.

Real-Time Image Analysis： AI algorithms integrated with imaging hardware enable real-time analysis, providing instant feedback during procedures. This technology is particularly valuable for guiding interventions or detecting anomalies in dynamic environments.

Explainable AI (XAI)： Explainable AI ensures transparency by providing visual or textual explanations for AI-generated decisions. Techniques such as heatmaps show focus areas in medical images, enhancing clinician trust and interpretability.

Benefits

AI automates repetitive tasks in the medical imaging process such as image analysis and report generation. This enhanced diagnostic efficiency allows healthcare professionals to dedicate more time to patient care (Pinto-Coelho, 2023). Along with speed, AI systems can analyze medical images with precision, helping detect malignancies that may be hard to identify with traditional methods by identifying subtle changes that may be overlooked by clinicians. For instance, AI algorithms can detect minute anomalies in radiological images, such as early-stage tumors, thereby reducing human error and facilitating prompt treatment.

Additionally, AI in medical imaging has been able to improve accessibility to medical resources in remote areas with multiple developments. AI algorithms analyze medical images on-site, providing immediate diagnostic insights. This is particularly beneficial in areas lacking specialists, as AI can assist in early disease detection and treatment planning (OpenMedScience, 2024). AI also enables the remote analysis of medical images so that healthcare professionals in urban centers can interpret images from rural patients, ensuring expert diagnostics without the need for patient relocation (Bansal, 2024). The smartphone-based AI applications can also provide benefits to users in remote areas by providing preliminary evaluations that can guide them to seek further medical attention if necessary (Higzi, 2024).

AI can also support personalized and precision medicine. For example, AI can be used to predict responses of tumors by incorporating a patient's genomic profile along with their imaging data (Bansal, 2024). Wearable devices and digital health tools also allow the integration of real-time data analysis and continuous monitoring (Minttihealth, 2023). This invites patients to actively manage their own care while also improving overall accuracy of their diagnosis. In the field of cardiology, AI can predict a patient's risk of stroke or future heart attacks based on their heart imaging data. Together, the predictability and data integration capabilities of AI can facilitate more proactive and personalized treatment plans (Bansal, 2024).

Regulatory and Ethical Issues

Regulatory Frameworks and Challenges：AI systems differ from traditional medical devices due to their ability to learn and evolve. Existing frameworks like the FDA's are designed for static devices producing consistent outputs. To comply, AI models are often "locked" before deployment, restricting adaptation. While this ensures initial safety, it limits AI's ability to improve, potentially reducing diagnostic accuracy and relevance over time.

Emerging regulatory approaches aim to address this limitation. Transparency in model development and updates is now a key requirement, ensuring stakeholders understand how AI evolves. Real-time monitoring mechanisms allow continuous oversight, enabling early issue detection and correction. Adaptive frameworks are being developed to permit safe updates post-deployment. These measures balance innovation with rigorous safety standards, ensuring AI systems remain reliable and effective.

Data Privacy and Ethical Considerations： AI's reliance on large datasets of sensitive patient information raises significant privacy concerns. Regulations such as GDPR and HIPAA mandate anonymization, informed consent, and transparent data usage to maintain trust and encourage adoption. Beyond compliance, robust safeguards are essential to prevent breaches and misuse, balancing the need for comprehensive datasets with protecting patient confidentiality. Transparent communication about data usage and AI-generated insights is critical to fostering trust.

Bias in training datasets poses another ethical challenge. Models trained on non-representative data, such as datasets predominantly from Caucasian patients, can produce suboptimal diagnostics for minority groups, worsening healthcare disparities. Developers must prioritize diverse, representative datasets and implement mechanisms to identify and flag uncertainty when encountering underrepresented groups. Addressing these biases is vital to ensure equitable healthcare outcomes.

Accountability and Integration： AI-driven diagnostic errors underscore the need for clear accountability frameworks to define the roles of developers, healthcare providers, and institutions. Trust in AI systems relies on their function as tools that enhance, rather than replace, physician expertise—proper integration positions AI as a supportive resource, augmenting clinical decision-making while preserving provider accountability.

Key Players

Companies/Hospitals/Research Institutes/Nonprofits

All of the tech companies mentioned throughout this chapter are key participants in the development of medical imaging in AI. These companies are at the forefront of its development, driving its innovation to improve diagnostic accuracy, efficiency, and patient outcomes. Many of these companies partner with healthcare professionals and researchers to ensure its widespread adoption and continuous improvement. For example, GE Healthcare collaborates with hospitals and research institutions like the Mayo Clinic and the Karolinska Institute in Sweden (Taylor, 2023). The chair of radiology at Mayo Clinic stated “this collaboration brings our research and clinical teams’ expertise and feedback closer to product development and commercialization of innovation, ultimately accelerating the rate of translation in our research to patient care and offering greater opportunity for global impact.” Alliances like these can accelerate the development of technologies and increase their impact, along with the company’s profits.

These companies leveraging AI in medical imaging are also working together to break the barriers in their acceptance (Aidoc, 2024a). As the Transformation Officer at Aidoc declared, “AI holds the potential to revolutionize patient care, but its progress is stalled by fragment systems and the inability to scale effectively.” Aidoc recently announced its collaboration with NVIDIA in pushing out the BRIDGE guideline, a Blueprint for Resilient Integration and Deployment of Guided Excellence, that aims to “accelerate AI adoption across the healthcare industry.”

Hospitals, research institutes, and nonprofits are heavily intertwined with these medical imaging companies. A press release from Paige in July of 2024 announced the integration of Paige’s AI technology with three UK Hospital systems. North Bristol NHS Trust’s Southmead Hospital, University Hospitals Coventry and Warwickshire NHS Trust, and Oxford University Hospitals NHS Foundation Trust are implementing the Paige Prostate Suite into their standard care to evaluate its effectiveness in diagnosing prostate cancer. This initiative, part of the ARTICULATE PRO study led by the University of Oxford, aims to assess how this AI application influences clinical decision-making, pathology service delivery, and resource utilization in real-world settings. The study is funded by the Accelerated Access Collaborative's Artificial Intelligence in Health and Care Award, overseen by the Department of Health and Social Care.

Nonprofit organizations like the Medical Image Computing and Computer Assisted Intervention (MICCAI) Society and the American College of Cardiology (ACC) are actively engaged in advancing AI applications in medical imaging through various initiatives. MICCAI has introduced models such as VISTA3D, a unified segmentation foundation model, and MAISI (Medical AI for Synthetic Imaging), a 3D diffusion model. Both are open-source, promoting accessibility and collaboration in medical imaging (MICCAI Society, 2024). The MICCAI Society has also developed the FAIMI (Fairness of AI in Medical Imaging) Workshop, which explores ethical considerations in deploying AI systems in medical imaging, aiming to ensure equitable and accountable AI applications (FAIMI Workshop, 2023).

The ACC has launched a resource center offering clinicians curated AI materials to facilitate the integration of AI into cardiovascular care, including educational content and updates on AI regulations (ACC, 2024). The ACC also partnered with Aidoc, with their VP on innovation expressing that “together, we can revolutionize how coronary artery calcium is detected and treated, profoundly impacting countless patients’ lives, and that “this collaboration is a testament to our shared commitment to pushing the boundaries of medical innovation” (Aidoc, 2024b).

Patients and Doctors

In an article that gathered information on patient perspective on AI in radiology, they found that patients generally support sharing deidentified data if it serves the common good, aids others with similar health conditions, or contributes to research (Borondy Kitts, 2023). However, concerns about privacy risks, data misuse, re-identification risks, and lack of transparency persist. In an anonymized survey, 75% of respondents trusted AI when its results were reviewed by a physician, but only a third trusted AI to work alone, demonstrating a lack of trust in unsupervised AI. Loss of personal interaction and empathy with healthcare professionals is also a significant concern with increased reliance on AI. However, this could be mitigated by AI’s ability to enhance the doctor-patient relationship by reducing routine workloads for physicians, freeing up time for personalized interactions.

Clinicians have diverse perspectives on AI in medical imaging, with opinions influenced by their experiences, expectations, and concerns about the technology. The previously mentioned press release from Paige quoted doctors with positive outlooks on the incorporation of AI into their patient care, with one stating that they have “studied the disease and progression of prostate cancer in clinical research for over 25 years, it is a significant advancement that Paige’s AI applications have achieved a level of validation and performance that allows safe and effective live clinical use” (Paige, 2024). They continued by adding that “using Paige Prostate Suite alongside our standard of care has the promise to increase efficiency and improve reproducibility of results for patients.”

However, Harvard Medical School Study that examined the variable impact of AI assistance on radiologists' diagnostic accuracy found that AI enhances performance for some and diminishes it for others, with these effects not consistently linked to factors like years of experience, specialization in thoracic radiology, or prior AI usage (Pesheva, 2024). The findings highlight the necessity for personalized integration of AI tools into clinical practice, moving away from a universal approach to ensure that AI serves as a beneficial aid rather than a hindrance. Additionally, the research team stressed the importance of training radiologists to detect inaccurate results, question AI’s diagnostic conclusions, and the cooperation between clinicians and AI developers to improve operability.

Case Studies

Detecting COVID-19 From Chest X-Rays and CT Scans

To aid in the rapid diagnosis and management during the 2020 COVID-19 outbreak in the U.S., AI models were trained on thousands of chest X-rays and CT scans to differentiate between COVID-19 pneumonia and other conditions (Fusco et al., 2021). The systems were able to achieve diagnostic accuracy comparable to expert radiologists and helped during peak pandemic times when healthcare systems were overwhelmed. Overall, this helped improve triage and resource allocation in hospitals, especially in remote areas.

Predicting Stroke Outcomes from Imaging

The e-Stroke Suite software by Brainomix is an AI-powered platform designed to assist in the diagnosis and treatment of strokes (Brainomix, 2024). The program uses AI algorithms to analyze CT scans to predict stroke outcomes and guide treatment, such as thrombectomy decisions. The system provides a clear map of damaged brain tissue and highlights areas that could benefit from intervention. The reduced decision-making time and increased access to effective stroke care offers worldwide benefits.

AI-Guided Tumor Ablation

Exablate Neuro is an AI-guided MRI-guided focused ultrasound system produced by Insightec. It is used to non-invasively ablate (destroy) tumors, including cancerous brain lesions (Insightec, 2024). The real-time AI algorithms can adjust the ultrasound waves to target tumors with precision, sparing healthy tissue. The utilization of this technology can lead to faster recovery times and fewer complications compared to traditional surgery. Insightec continues to collaborate and innovate, with researchers evaluating their technology in a range of clinical trials to expand its applications and improve patient outcomes.

Future Contributions

Some sections are blank to be left open for future contributors. Use these as a starting point for adding onto the existing work.

Challenges and Limitations:

• privacy and ethics
• data quality and labeling
• interpretability
• generalization and biases

Current trends and Future Directions:

• multimodal AI and quantum computing
• federated learning for protecting data privacy
• real time analysis during surgery
• social and political implications of widespread AI adoption in healthcare

Notable Datasets

• MIMIC-CXR
• NIH ChestX-ray14
• CheXpert

References

ACC. (2024, April 22). Access ACC’s new AI Resource Center. https://www.acc.org/Membership/Sections-and-Councils/Health-Care-Innovation-Section/Section-Updates/2024/04/22/18/40/AI-Resource-Center

Ai in radiology: Pros & cons, applications, and 4 examples. V7. (2022, November). https://www.v7labs.com/blog/ai-in-radiology?utm_source

AI Medical Imaging Annotation: Healthcare data labeling. V7. (n.d.). https://www.v7labs.com/darwin/medical-imaging-annotation

Aidoc. (2024a, October 1). Aidoc to establish guideline to accelerate AI adoption in healthcare in collaboration with NVIDIA. PR Newswire. https://www.prnewswire.com/il/news-releases/aidoc-to-establish-guideline-to-accelerate-ai-adoption-in-healthcare-in-collaboration-with-nvidia-302280843.html

Aidoc. (2024b, February 29). Aidoc and American College of Cardiology collaborate to revolutionize cardiovascular care with best-in-class AI. PR Newswire. https://www.prnewswire.com/news-releases/aidoc-and-american-college-of-cardiology-collaborate-to-revolutionize-cardiovascular-care-with-best-in-class-ai-302074988.html

Bansal, N., MD, MS-HSM. (2024). AI in medical imaging: Transforming healthcare diagnostics. Following Healthcare. https://followinghealthcare.com/ai-in-medical-imaging/

Birnhak, S. (2024, August 1). How ai is revolutionizing medical diagnosis for enhanced accuracy. RSR. https://readysetrecover.com/blog/ai-and-medical-diagnosis

Borondy Kitts, A. (2023). Patient perspectives on artificial intelligence in radiology. Journal of the American College of Radiology, 20(9), 863-867. https://doi.org/10.1016/j.jacr.2023.05.017

Brainomix. (2024). e-Stroke Suite: Revolutionizing stroke care with AI. Brainomix. https://www.brainomix.com/stroke/

Densford, F. (2019, January 11). Image analysis dev Koios Medical raises $5m. MassDevice. https://www.massdevice.com/image-analysis-dev-koios-medical-raises-5m/

FAIMI Workshop. (2023, October). Fairness of AI in Medical Imaging (FAIMI) Workshop at MICCAI 2023. FAIMI. https://faimi-workshop.github.io/2023-miccai/

FDA. (2018, April 11). FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. U.S. Food and Drug Administration. https://www.fda.gov/news-events/press-announcements/fda-permits-marketing-artificial-intelligence-based-device-detect-certain-diabetes-related-eye

Fukushima, R. (2023, February 10). Welcome Arterys: Integrating medical imaging insights to drive better outcomes for patients. Tempus. https://www.tempus.com/resources/content/blog/welcome-arterys-integrating-medical-imaging-insights-to-drive-better-outcomes-for-patients/

Fusco, R., Grassi, R., Granata, V., Setola, S. V., Grassi, F., Cozzi, D., Pecori, B., Izzo, F., & Petrillo, A. (2021, September 30). Artificial intelligence and COVID-19 using chest CT scan and chest X-ray images: Machine learning and deep learning approaches for diagnosis and treatment. J Pers Med, 11(10), 993. https://doi.org/10.3390/jpm11100993

Higzi, R. (2024). Leveraging AI for improved healthcare accessibility. Klarity Health. https://my.klarity.health/leveraging-ai-for-improved-healthcare-accessibility/

Insightec. (2024). About us. Insightec. https://insightec.com/about-us/

MICCAI Society. (2024, November 13). MICCAI industrial talk: Open-source foundation models for 3D medical image segmentation and generation. MICCAI. https://miccai.org/index.php/news/2024/11/13/miccai-industrial-talk-open-source-foundation-models-for-3d-medical-image-segmentation-and-generatio

Minttihealth. (2023, December 6). Integrating artificial intelligence with wearable devices and real-time monitoring for enhanced healthcare outcomes. Minttihealth. https://minttihealth.com/integrating-artificial-intelligence-with-wearable-devices-and-real-time-monitoring-for-enhanced-healthcare-outcomes/

OpenMedScience. (2024). The transformative role of AI in MRI: Enhancing diagnostics, reducing costs, and improving patient access. OpenMedScience. https://openmedscience.com/the-transformative-role-of-ai-in-mri-enhancing-diagnostics-reducing-costs-and-improving-patient-access/

Paige. (2024). Paige AI: Transforming cancer diagnosis with AI. Retrieved November 27, 2024, from https://paige.ai/

Paige. (2024, July 22). Three UK-based hospital systems exploring the power of Paige’s AI for diagnosing prostate cancer in live clinical settings. Paige. https://paige.ai/three-uk-based-hospital-systems-exploring-the-power-of-paiges-ai-for-diagnosing-prostate-cancer-in-live-clinical-settings/

Pesheva, E. (2024, March 19). Does AI help or hurt human radiologists’ performance? It depends on the doctor. Harvard Medical School. https://hms.harvard.edu/news/does-ai-help-or-hurt-human-radiologists-performance-depends-doctor

Pinto-Coelho, L. (2023). How artificial intelligence is shaping medical imaging technology: A survey of innovations and applications. Bioengineering, 10(12), 1435. https://doi.org/10.3390/bioengineering10121435

Saidy, N. T. (2021). Artificial intelligence in healthcare: opportunities and challenges. YouTube. https://www.youtube.com/watch?v=uvqDTbusdUU

Savoy, M. (2020). IDx-DR for diabetic retinopathy screening. American Family Physician, 101(5), 307–308. https://www.aafp.org/pubs/afp/issues/2020/0301/p307.html

SkinVision. (2024). SkinVision: Skin cancer melanoma detection app. Retrieved November 27, 2024, from https://www.skinvision.com/

Taylor, N. P. (2023, September 15). GE HealthCare partners with Mayo Clinic to advance medical imaging and theranostics. MedTech Dive. https://www.medtechdive.com/news/ge-healthcare-partner-mayo-clinic/693780/

Ulve, K. (n.d.). Ai’s Ascendance in Medicine: A timeline. Cedars. https://www.cedars-sinai.org/discoveries/ai-ascendance-in-medicine.html