Revolutionizing healthcare: AI’s impact on medical imaging

Medical imaging is undergoing a new era of artificial intelligence, revolutionizing healthcare and diagnosis treatment.

Source: Wikimedia Common

What is medical imaging? According to the National Institute of Health (NIH), medical imaging is the visual representation of different tissues and organs of the human body to monitor the body's normal and abnormal anatomy and physiology (Shah, 2022). Some standard techniques include Computed Tomography (CT), Positron Emission Tomography (PET), and Magnetic Resonance Imaging (MRI).

Accurate medical scans are essential for assessing a patient’s disease and providing effective treatment. Nevertheless, traditional image interpretation methods can be time-consuming and subject to human error. With the number of CT scans per thousand population increasing by 30% over the past decade and 12% for MRIs, radiologists have had a profusion of data to analyze, drastically increasing their workload (The Canadian Medical Imaging Inventory, 2023). This is where AI steps in; AI algorithms can process and analyze images much faster, significantly reducing the time taken to diagnose a patient (Khalifa, 2024). This speed is pivotal in emergencies where every second matters. AI also enhances the accuracy of diagnoses. By learning from large datasets of medical images, AI algorithms can identify patterns and anomalies that the human eye might overlook. This increased accuracy is vital in reducing misdiagnoses and ensuring patients receive the most effective treatment.

Applications of AI in Medical Imaging

Artificial Intelligence is being increasingly used in medical imaging for the early detection of tumors, cancers, cardiovascular diseases, and neurological disorders such as Alzheimer’s and Parkinson’s disease (Najjar, 2023). A notable application of AI technology is the generation of synthetic CT scans (sCT) from MRI’s. In cases where a brain tumor is suspected, an MRI scan is typically performed to identify any irregularities. AI algorithms can convert the original MRI scan into a synthetic CT scan rather than subjecting the patient to a potentially unnecessary CT scan for further evaluation. This approach not only conserves time and financial resources for patients but also minimizes their exposure to the high doses of radiation associated with conventional CT scans.

Additionally, the alignment and combination of CT and MRI scans, known as "registration," is often required to analyze the tumor's characteristics comprehensively. When executed manually by radiologists, this process can be labor-intensive. However, by employing AI to transform the MRI scan into an sCT scan, these images can be automatically aligned and combined, facilitating the analysis by radiologists and enhancing the accuracy of treatment planning. To prove their accuracy and reliability, a study in 2021 compared the difference between CT scans and sCT scans in a group of 20 patients with brain malignancies (Paudyal, 2023). These results showed that the sCT and CT images were comparable in dosimetric and geometric evaluation. Dosimetric evaluations measure the amount and distribution of radiation delivered to a tumor, preventing patients from being overexposed to toxic doses of radiation (Zhong, 2023). In contrast, geometric evaluation refers to the proper alignment of CT and sCT scans as well as the accurate targeting of tumors. Thus, the study successfully validated a commercially available CNN-based software for sCT utilization.

Types of AI used in Medical Imaging

Several AI techniques are employed in medical imaging, with supervised learning being one of the most prominent. In supervised learning, models are trained on large datasets of labeled medical images to recognize patterns and predict outcomes, allowing them to grow more accurately over time (Pinto-Coelho, 2023). This method enables AI systems to identify diseases by comparing new images to previously analyzed ones. For instance, models trained on thousands of images showing various stages of cancer can learn to detect similar patterns in new images, aiding in diagnosing patients faster and more accurately (Jader, 2022).

While supervised learning relies on labeled datasets, deep learning models like CNNs and GANs are engineered to learn automatically and extract features from data. CNNs, a deep learning algorithm, excel at analyzing large datasets by learning hierarchical features from images (IBM, 2024). This algorithm uses a process that involves three key stages: training, optimization, and inference. During training, the model is fed labeled data and learns to recognize specific features, such as tumors in an MRI scan. In the optimization stage, the model adjusts its internal parameters to improve accuracy. Finally, during inference, the model applies what it has learned to new data, making predictions in real time.

Various studies demonstrate the success of CNNs in medical imaging. For instance, research shows that CNNs can outperform traditional methods in detecting lesions and characterization tasks such as the lesion's size, shape, and texture. A study found that CNNs outperform the human observer, particularly at high noise levels where it can be difficult to distinguish subtle gradations of light and shadow (De Man, 2019). This level of precision can significantly improve patient outcomes by enabling earlier detection and treatment.

GANs, another deep learning architecture, have also gained traction in medical imaging. They consist of two competing neural networks: a generator that creates synthetic data, and a discriminator that evaluates the data's authenticity (Arora, 2022). In radiology, these GANs can generate new, high-quality images from existing data, such as enhancing the resolution of medical scans or creating synthetic images to expand training datasets. These powerful AI tools have shown promise in generating high-quality synthetic images that can be used to train machine-learning models for tasks including medical diagnosis and treatment planning.

During a study to consider how realistic GANs’ synthetic images were, 4 expert radiologists with over 15 years of experience were asked to determine between a generated MRI scan and a real one (Skandarani, 2023). Each expert was shown 100 images, 50 of which were synthetically generated and 50 of which were real scans. The results showed that the experts were right only 60% ± 10% of the time, highlighting the potential of GANs in medical imaging.

Technical Challenges

Despite these advancements, several technical challenges persist in integrating AI into medical imaging. One notable hurdle is the cost associated with acquiring and labeling high-quality images. Each MRI or CT scan can range from $500 to $3,000, making it financially burdensome for healthcare institutions to develop extensive training datasets (CT Scan vs MRI, 2023). Additionally, the images need to be labeled by expert radiologists to train the model, which can be costly and time-consuming. Lastly, many inconsistencies in imaging protocols across institutions can affect the generalizability of AI models (Eche, 2021). Nevertheless, following interoperability protocols such as FHIR, emphasizing interinstitutional collaborations, and including explainability tools to provide transparency into how algorithms reach their decision can be important steps in enhancing the generalizability of AI models (Theriault-Lauzier, 2024).

Ethical Considerations

The technical considerations of using AI in medical imaging are closely tied to ethical and security concerns. While AI promises improved diagnostic accuracy and efficiency, it heavily depends on large volumes of patient data, raising significant privacy issues (Herrington, 2023). Data sharing between institutions is crucial to enhance AI performance however this increases the risk of confidentiality breaches, especially as over 5,000 data breaches have occurred in the last decade (Alder, 2024). Moreover, the possibility of replacing healthcare professionals with AI tools introduces ethical dilemmas, as many patients still prefer human oversight in their care. Additionally, the security of AI-generated images is a growing concern. A recent study demonstrated that hackers could tamper with CT scans to artificially add or remove lung cancer, with a 99.2% success rate for cancer injection and 95.8% for removal when radiologists were unaware of the manipulation (Chu, 2020). Even when warned, the success rate for cancer removal remained high, illustrating the potential for sophisticated attacks. If such breaches infiltrate radiology systems, it could undermine clinical workflows, as radiologists would need to verify the authenticity of every image, adding significant strain to an already demanding process.

Business Implications of AI in Medical Imaging

Integrating AI into medical imaging transforms healthcare and presents significant business opportunities and challenges for stakeholders across the industry. For healthcare providers, AI-driven solutions can save costs by streamlining the diagnostic process and reducing the need for repeated scans. However, the upfront investment in AI technologies, such as acquiring and labeling data, can be financially burdensome, requiring institutions to evaluate long-term ROI. Data breaches, which have increased by 240% from 2018 to 2023 signal a call for policymakers to establish stricter regulations and frameworks around data privacy. This push for tighter security measures will likely lead to increased spending on cybersecurity, creating opportunities for tech firms specializing in data protection. Simultaneously, the insurance industry must adapt, as the rise of AI-generated diagnoses may necessitate new policies regarding liability, especially in cases of misdiagnosis or breaches. Stakeholders, from hospital administrators to policymakers and investors, must collaborate to ensure that AI is deployed ethically and effectively, guaranteeing technological innovation and patient safety.

Conclusion

In conclusion, AI is revolutionizing medical imaging by increasing diagnostic accuracy, speeding up the analysis of medical scans, and reducing patient exposure to radiation. Its applications in generating synthetic scans, detecting early diseases, and even improving the quality of medical images demonstrate its transformative potential. However, the technical and ethical challenges, including the cost of data acquisition, labeling, and privacy concerns, must be addressed. While AI has the potential to reduce healthcare costs and improve patient outcomes, it also raises new questions for business leaders and policymakers regarding data security and the role of healthcare professionals. As AI evolves, balancing innovation with ethical considerations will be essential to successfully integrating it into healthcare.

References:

Alder, Steve. Healthcare Data Breach Statistics, 24 Sept. 2024, www.hipaajournal.com/healthcare-data-breach-statistics/. Accessed 21 Oct. 2024.

Arora, Anmol, and Ananya Arora. “Generative Adversarial Networks and Synthetic Patient Data: Current Challenges and Future Perspectives.” Future Healthcare Journal, U.S. National Library of Medicine, July 2022, pmc.ncbi.nlm.nih.gov/articles/PMC9345230/. Accessed 20 Oct. 2024.

The Canadian Medical Imaging Inventory: 2022–2023, www.cda-amc.ca/sites/default/files/hta-he/HC0024_cmii_2022_2023_evidence_preview.pdf. Accessed 21 Oct. 2024.

Chu, Linda C, et al. “The Potential Dangers of Artificial Intelligence for Radiology and Radiologists.” Journal of the American College of Radiology : JACR, U.S. National Library of Medicine, Oct. 2020, pmc.ncbi.nlm.nih.gov/articles/PMC7164850/. Accessed 20 Oct. 2024.

“CT Scan vs. MRI - Uses, Cost, Risks & More.” Home, 12 Oct. 2023, www.docpanel.com/ct-scan-vs-mri-uses-cost-risks-more/. Accessed 20 Oct. 2024.

De Man, Ruben, et al. “Comparison of Deep Learning and Human Observer Performance for Detection and Characterization of Simulated Lesions.” Journal of Medical Imaging (Bellingham, Wash.), U.S. National Library of Medicine, Apr. 2019, pmc.ncbi.nlm.nih.gov/articles/PMC6586983/. Accessed 20 Oct. 2024.

Eche, Thomas, et al. “Toward Generalizability in the Deployment of Artificial Intelligence in Radiology: Role of Computation Stress Testing to Overcome Underspecification.” Radiology. Artificial Intelligence, U.S. National Library of Medicine, 27 Oct. 2021, pmc.ncbi.nlm.nih.gov/articles/PMC8637230/. Accessed 20 Oct. 2024.

Herington, Jonathan, et al. “Ethical Considerations for Artificial Intelligence in Medical Imaging: Data Collection, Development, and Evaluation.” Journal of Nuclear Medicine : Official Publication, Society of Nuclear Medicine, U.S. National Library of Medicine, 1 Dec. 2023, pmc.ncbi.nlm.nih.gov/articles/PMC10690124/. Accessed 20 Oct. 2024.

Hussain, Shah, et al. “Modern Diagnostic Imaging Technique Applications and Risk Factors in the Medical Field: A Review.” BioMed Research International, U.S. National Library of Medicine, 6 June 2022, pmc.ncbi.nlm.nih.gov/articles/PMC9192206/. Accessed 20 Oct. 2024.

IBM. “What Are Convolutional Neural Networks?” IBM, 2 Oct. 2024, www.ibm.com/topics/convolutional-neural-networks. Accessed 20 Oct. 2024.

Jaber, Nadia. “Can Artificial Intelligence Help See Cancer in New Ways?” Can Artificial Intelligence Help See Cancer in New Ways? - NCI, 22 Mar. 2022, www.cancer.gov/news-events/cancer-currents-blog/2022/artificial-intelligence-cancer-imaging. Accessed 20 Oct. 2024.

Khalifa, Mohamed. “AI in Diagnostic Imaging: Revolutionising Accuracy and Efficiency.” Computer Methods and Programs in Biomedicine Update, Elsevier, 5 Mar. 2024, www.sciencedirect.com/science/article/pii/S2666990024000132#:~:text=Traditional%20methods%20of%20image%20interpretation,counts%20%5B2%2C3%5D. Accessed 20 Oct. 2024.

Najjar, Reabal. “Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging.” Diagnostics (Basel, Switzerland), U.S. National Library of Medicine, 25 Aug. 2023, pmc.ncbi.nlm.nih.gov/articles/PMC10487271/#:~:text=AI%20also%20plays%20a%20critical,patterns%20associated%20with%20these%20conditions. Accessed 20 Oct. 2024.

Paudyal, Ramesh, et al. “Artificial Intelligence in CT and MR Imaging for Oncological Applications.” Cancers, U.S. National Library of Medicine, 30 Apr. 2023, pmc.ncbi.nlm.nih.gov/articles/PMC10177423/. Accessed 20 Oct. 2024.

Pinto-Coelho, Luís. “How Artificial Intelligence Is Shaping Medical Imaging Technology: A Survey of Innovations and Applications.” Bioengineering (Basel, Switzerland), U.S. National Library of Medicine, 18 Dec. 2023, pmc.ncbi.nlm.nih.gov/articles/PMC10740686/. Accessed 20 Oct. 2024.

Skandarani, Youssef, et al. “Gans for Medical Image Synthesis: An Empirical Study.” Journal of Imaging, U.S. National Library of Medicine, 16 Mar. 2023, pmc.ncbi.nlm.nih.gov/articles/PMC10055771/. Accessed 20 Oct. 2024.

Theriault-Lauzier, Pascal. “A Responsible Framework for Applying Artificial Intelligence on Medical Images and Signals at the Point of Care: The PACS-Ai Platform.” Canadian Journal of Cardiology, Elsevier, 15 June 2024, www.sciencedirect.com/science/article/pii/S0828282X24004276. Accessed 20 Oct. 2024.

Zhong, Yang, et al. “Geometric and Dosimetric Evaluation of Deep Learning Based Auto-Segmentation for Clinical Target Volume on Breast Cancer.” Journal of Applied Clinical Medical Physics, U.S. National Library of Medicine, July 2023, pmc.ncbi.nlm.nih.gov/articles/PMC10338811/. Accessed 20 Oct. 2024.