AI in medical imaging: Pushing the boundaries of precision diagnosis
8 MINS READ
Improving healthcare with AI-led diagnosis
Innovation in medical imaging is gaining pace. Medical images are a rich source of patient data, but can be complex to interpret, taking up valuable time and resources. Artificial intelligence (AI) can analyze medical images quickly and precisely, bolstering precision medicine and personalized healthcare. However, challenges and uncertainties surround the application of AI in medical imaging solutions. Privacy and security of patient data shared across digital channels, lack of clarity on accuracy of AI’s outcome and data bias can impact the reliability of diagnosis. Despite these challenges, the future of AI in medical imaging looks promising, capable of easing workload for radiologists, and further reducing variables and outliers that take the subjectivity out of diagnosis.
AI aids decision-making across businesses and in our everyday lives. So, what makes an AI-driven decision superior to human cognition? It is fast and takes guesswork out of decision-making.
In healthcare, this could translate to improved patient experience, and even save lives through preventive and personalized care. In recent years, AI is gaining momentum in medical imaging for a wide array of clinical conditions and processes. Advances in preoperative planning, robotics, and navigation which use imaging in some form for visualization and execution have opened-up avenues for AI-based technologies. Researchers at Tulane University discovered that AI can accurately detect and diagnose colorectal cancer by analyzing tissue scans, in some cases better than pathologists.
Despite its ability to reduce variables and outliers for improved clinical workflows, diagnosis, and patient outcomes, AI applications have inherent challenges such as algorithmic bias and uncertainty in prediction. Moreover, it may take time for clinicians to gain confidence in AI supported diagnosis. There are also privacy and security concerns when medical images are shared through digital channels. With the explosion of data due to digitization, many radiologists also doubt the ability of AI to interpret many images consistently without error. Even though these challenges exist, medical imaging is one of the most promising areas for research in the application of AI. This paper describes ways to overcome the uncertainties surrounding AI-based diagnostic solutions and how they can be leveraged to aid clinical decision-making for improved and personalized healthcare delivery.
AI-driven diagnosis and its benefits
Medical imaging such as X-rays, Ultrasound, CT, PET, and MRI are important tools for diagnosis, treatment planning, and long-term health monitoring. However, analyzing these medical images manually requires significant expertise, time, and effort. In addition, increased workload for radiologists and physicians may cause potential delays in disease diagnosis and treatment.
The performance of AI models has surpassed human intelligence in other areas such as surveillance and autonomous driving, achieving high levels of accuracies. With its ability to analyze enormous volumes of image data, AI can have an even more profound impact on healthcare operational efficiency and patient outcomes through faster, more accurate diagnosis and early intervention. Some of the areas in healthcare where AI is being used are:
Auto-detection of specific health conditions: Advanced AI techniques applied to digitized imaging data helps in the automatic detection of health conditions. For example, it can identify anomalies from retinal images, bleeding spots in gastrointestinal (GI) tract as well as lesions, and COVID-19 severity scores from computed tomography (CT) scans. AI in medical imaging also aids in tracking the progression of wound healing, detecting and categorizing portal hypertension, and hepatobiliary disorders, as well as screening of cancer from a variety of imaging modalities.
Auto-identification of organs and their substructures: There are emerging AI techniques that can automatically identify cardiac sub-structures from CT angiography, anatomical landmarks from X-rays, and human cells in digital slides.
Synthesizing multiple image modes:
Multimodal MRI imaging: Complementary information obtained from different contrasts of tissues helps physicians diagnose diseases more accurately and plan treatments effectively. However, acquiring multiple contrasts MRI for every patient is not time-efficient and cost-effective. Hence, AI-based synthesis of multimodal MRI is an effective solution.
Time series analysis of images taken over time: Time-series analysis of medical images facilitates the deeper understanding of a patient’s entire lifetime’s health and clinical trajectories, across a wide range of diseases, including cancer, and cardiovascular diseases. It is an important tool in medical imaging for accomplishing tasks such as dynamic forecasting, survival analysis, early diagnosis, and treatment affect estimation.
Assisting surgeons: Several AI-based solutions are being developed to support surgery preparations and serve as building blocks for remote surgery. These include auto-instrument positioning during eye surgeries, tool detection and identification, smoke detection and dehazing, and surgical visualization.
As a result of the ongoing developments in medical imaging techniques, there are countless avenues for developing automated diagnostic solutions using these medical images.
AI-based diagnosis can prevent unnecessary re-admissions and invasive testing, thus leading to reduction in waiting time, patient anxiety, and healthcare costs.
From precision medicine and personalized healthcare to risk-prediction of diseases and detection of abnormalities, AI-based diagnosis can enhance clinical workflows, improve patient experience, and elevate overall care. Furthermore, it prevents unnecessary re-admissions and the need for additional testing, leading to reduction in waiting time, patient anxiety, and healthcare costs. Automation accelerates decisions related to diagnosis and treatment by ensuring that images and reports are more easily accessible through electronic health records (EHRs) or through text messages on mobile devices.
Trust and transparency: The biggest barriers to AI-adoption in clinical practice
The 2020 survey conducted by the American College of Radiology (ACR) Data Science Institute revealed that 30% of radiologists are currently using AI as part of their practice; 20% plan to purchase AI tools in the next 1 to 5 years.
With its ability to analyze images quickly and precisely, AI has proven to be a valuable companion for radiologists, pathologists, and physicians for faster diagnosis, risk assessment, and treatment. Although automation in medical imaging has progressed over the past few years, it continues to face resistance in clinical settings. Increase in adoption will depend on addressing the two main barriers which revolve around trust and transparency.
In recent years, data sharing between hospitals and AI companies has highlighted several ethical questions surrounding patient privacy, ethics of data ownership, accuracy, and the consequences of a security breach. As a result of increasing regulatory requirements for data protection, healthcare organizations need to take proactive measures to implement security and privacy measures.
The solution: Simple steps such as internal staff training, data encryption, and a heightened focus on connected device risks can help protect patient data. Blockchain technology is another viable option. It not only serves as an efficient data sharing platform for medical imaging and patient history, but also as a solution to prevent data breaches.
Another challenge with AI-based solutions is the inherent data bias, uncertainty, and reliability of its diagnosis. This is primarily due to its reliance on the quality of annotated data, which could be erroneous. The availability of labeled data is limited since annotation relies on human expertise and manual intervention. To circumvent this issue, new AI techniques such as few-shot learning, semi-supervised, and self-supervised learning are being incorporated in developing automated diagnosis solutions.
AI models are generally biased towards the distribution of data used for its training and fail to generalize when applied to datasets coming from different hospitals, scanners, and acquisition protocols. Due to the black-box nature of deep models, it is difficult to gain the trust of doctors and patients to use these automated solutions in a clinical setting. Doctors find it difficult to decide when to rely on automated diagnoses and when not to.
The solution: There is a need to develop AI solutions that are generalizable and easily adaptable across different data domains. These solutions should provide outputs with an associated confidence score along with clinically relevant visuals and textual explanations. This will help physicians decide whether further tests are necessary to make a correct diagnosis in case the confidence score is very low.
The future of medical imaging
Medical images are a valuable source of patient data but can also be increasingly complex to interpret with precision. The adoption of software-based value-added services driven by IoT and data analytics is likely to automate and make this process simpler and more accurate. For instance, it can support computer-aided diagnosis, large-scale screening programs, and data integration platforms which provide a unified view of the patient's health. Soon, every medical imaging machine will be connected to the cloud where innovative algorithms analyze data and help doctors screen, assess, and diagnose patients.
Assessments by the algorithm are more reproducible and less subjective. Studies have shown that AI algorithms are more effective and efficient in identifying diseases than human experts. AI algorithms can read images in less time with higher accuracy (up to 95% in some cases). For example, emerging research suggests that AI could play an important role in improving breast cancer screening outcomes. An AI model reduced false positive results by 25% and reduced radiologist workload by more than 62%. By analyzing vast amounts of historic patient data, AI solutions can also provide real-time support to clinicians to help identify risks. For example, hospital re-admission risks which highlight a patient’s increased chance of returning to the hospital within 30 days of discharge.
There is still a long way to go before healthcare organizations adopt AI for autonomous diagnosis. But the future looks promising. The success of AI in medical imaging will depend on education and training, to address the trust and reliability factors. Radiologists, pathologists, and physicians must understand how to integrate AI into the clinical practice and evaluate its efficacy, and when to apply the appropriate human intervention to get the most precise diagnosis.