According to Grand View Research, the global AI in the medical imaging market is expected to have a 34.8% compound annual growth rate from 2023 to 2030. There are three key reasons why investors from all over the world intend to invest in this technology:
- Earlier lesion detection
- More thorough patient tracking
- Identification of abnormalities unseen by the human eye
Now developers and radiologists are cooperating to create and train algorithms that are accurate enough for clinical use. In the US, every AI algorithm requires the approval of the U.S. Food and Drug Administration (FDA) before implementation. By 2022, FDA has approved 521 AI solutions, 75% of which belong to radiology, a field that requires extreme accuracy. Actually, some AI-powered software solutions in radiology already boast over 90% accuracy, which got our team's attention.
Mind Studios has been developing custom healthcare solutions for years, and the recent leap in the development of AI technology, as well as the benefits it can bring to the industry, have become a primary focus of our interest. And in this article, we are going to reveal the main applications, benefits, and challenges of AI implementation for medical imaging.
AI techniques used in medical imaging
Among all the AI techniques, the most commonly used in the medical imaging industry are machine learning, deep learning, and computer vision. Let’s take a look at what exactly each of them brings to radiology.
Machine learning is one of the technologies for artificial intelligence, and it aims to imitate the process of human learning with the use of algorithms and data to train from. In medical imaging, machine learning is used in training the AI on medical scans and other images to grant it the ability of pattern recognition. With this ability, the model becomes capable of making predictions and classifications of diagnoses based on the provided image.
Deep learning is a subset of machine learning, representing a neural network able to simulate brain behavior. This ability grants it to also simulate learning processes from large volumes of data. A neural network usually consists of no less than three layers, and with only one layer it can already be used for predictions. The other layers are used to improve the accuracy and thoroughness of such predictions, making the neural network more optimized. The deep learning technique in medical imaging is expected to assist radiologists in providing more accurate diagnoses and analyzing suspicious lesions.
Computer vision is another technology for artificial intelligence that makes it capable of processing both 2D and 3D images, videos, and models. In medical imaging, computer vision enables the AI to recognize and identify known issues in the provided medical scan, as well as analyze and differentiate between them.
Applications of AI in medical imaging
Today, there is no actual AI yet that can completely replace a radiologist, but is there a way of implementing AI to enhance the existing processes? In this section, we are going to talk about practical applications for artificial intelligence in medical imaging, and how it can play the role of a “second reader”.
Traditionally, classifying a brain tumor means surgery, where a doctor removes as much of the tumor’s mass as possible to analyze the sample thoroughly. It takes time for the sample to be processed, stained, and delivered to a pathologist for analysis. Although AI is incapable of replacing the operation stage, it can enhance the sample analysis and save some precious time for the patient.
For example – DeepGlioma AI was able to classify the tumors of 153 patients into correct subgroups with an accuracy of 93.3%. DeepGlioma neural network algorithms analyze brain tumor tissue scans and classify them based on the World Health Organization (WHO) classification system. Along with great accuracy, this AI imaging solution also provides results in under 90 seconds.
Recent researches show that machine-learning models can only slightly assist doctors in breast cancer screening and hence are still far from widespread implementation.
For example, a trained AI model was able to increase the diagnostics accuracy for ultrasound examinations of breast cancer. In a study, 10 different radiologists diagnosed 663 breast exams without AI’s help and had an average accuracy of 92%. Then, the assistance of the AI tool for medical imaging helped to increase the overall accuracy in those examinations to 96%.
However, for a more advanced type of screening — mammography, the AI showed worse accuracy than the decisions of practicing professionals. A combination of three studies involving more than 79,900 screened women compared 36 AI tools with the clinical decisions of radiologists. 34 out of 36 solutions were less accurate than a decision of a single specialist, and every single one of them performed worse than the consensual decision of two radiologists.
These studies lead to the conclusion that it’s too early to adopt artificial intelligence for clinical use in an advanced type of screening, but for more common ones it can already be of assistance.
As the main source of medical imaging in detecting neurological diseases is MRI scans of the brain, it takes time to analyze the results and find any vital changes that signal a developing condition. Unfortunately, human sight is imperfect, and when it comes to minor differences, even an experienced specialist may fail to notice them. Artificial intelligence can add some precision to classifying these scans, leaving fewer signals unnoticed.
To study the impact of AI-powered tools in neurological screening, the researchers allocated 656 brain features across the 115 regions of a brain that flag pathologies related to Alzheimer’s disease (AD). Trained on this segmentation, the machine-learning solution required only one brain scan on an MRI 1.5 Tesla model, to identify and predict the development of AD at early stages. The core algorithms of the model are already-known ones for classifying cancer tumors, adapted for working with a brain only.
Eventually, this model, trained on MRI scans from the Alzheimer’s Disease Neuroimaging Initiative, was able to correctly differentiate brain scans of patients with AD in 98% of cases.
Artificial intelligence can be useful in detecting heart diseases on cardiovascular MRI scans. Research shows that a machine learning algorithm trained on 1923 CMR scans from 13 different institutions was used in segmenting left ventricular blood volume and myocardium. Not only the fact of usage claims our attention, but also the statistics, showing only 1 in 479 cases of mis-segmentation.
This great precision of the algorithm also takes about 20 seconds per decision, greatly surpassing manual diagnostics, which take around 13 minutes on average. This shows that with greater training and a bigger number of scans to train from, AI can be both efficient and time-saving in detecting cardiovascular issues early.
Radiation dosage reduction
Medical imaging by means of MRI or CT inevitably leads to patients being exposed to radiation. As every exposure comes with a harmful effect on the human body, the ways to reduce it are always a challenging topic. Especially when it comes to pediatrics, reducing the dose of radiation in medical imaging can become vital for a growing body. And artificial intelligence has something to offer here.
The solution lies in AI’s deep learning image reconstruction (DLIR). The main advantage of using DLIR in radiology is that it can reconstruct images from lower-dose computed tomography. The main challenge is whether the image is readable enough when reducing the radiation doses and applying the image reconstruction.
And according to a study, convolutional neural networks (CNN) — the most common AI technique and architecture — can reduce radiation dosage by 36–70% without losing any diagnostic information. This study shows that artificial intelligence can play a crucial role in radiation dosage reduction, however, it still requires more thorough research, as only 3 possible imaging modalities (CT, PET/MRI, and mobile radiography) were covered.
As X-rays and CT are still the most common ways of finding any fractures and injuries, they leave space for human errors even when looked at with the eyes of a professional. Artificial intelligence can’t boast a much greater accuracy, as it also makes mistakes, however, it can aid a specialist in diagnosing more precisely and in a shorter amount of time.
In a recent study, an AI-powered tool trained on more than 55,000 scans showed a sensitivity of 92% when detecting fractures. Compared to the sensitivity of 91% that clinicians managed to demonstrate in that study, it is safe to say that AI is coming close to serving as a diagnostic assistant for musculoskeletal injuries.
Benefits and challenges of AI in medical imaging
As artificial intelligence is growing in abilities rapidly, it’s easy to predict its popularity in medical imaging in the nearest future. However, it’s important to understand the benefits of its adoption right now and not overlook the challenges it brings. To sum up everything we've mentioned above, we believe the most important benefits of medical imaging AI are:
- More accurate classifications. Enhanced with deep learning and computer vision algorithms, artificial intelligence is able to recognize even the smallest abnormalities in the provided scans regardless of their location. With greater pattern recognition abilities and large volumes of training data, AI can accurately classify the lesion it identifies.
- Enhanced analysis. With proper training, AI models are capable of providing detailed information about the recognized lesions, usually serving as a second opinion for radiologists in practice or leading to a point hidden from a naked eye of a specialist.
- Creation of 3D models. Artificial intelligence is not only able to analyze 3D models thanks to computer vision but can also assist in their creation. For a more thorough medical imaging analysis, AI can 3D visualize a part of a patient’s body to provide better diagnostics. For example, there is an AI that 3D models the coronary arteries of a patient based on their coronary CT scans.
- Faster diagnosis. Human doctors take time to carefully read the imaging, it's part of the diagnosing process. For example, in cardiology, radiologists spend around 13 minutes diagnosing a patient, and with AI, they are able to do it in 20 seconds.
When it comes to the challenges of AI in medical imaging, the crucial one is definitely privacy issues. To train a model properly, medical researchers need to access an existing dataset from patients’ medical and personal records. This becomes a big stumbling block for institutions, as there are many patients who are against participating in a model’s training with their personal data, and have the right to do so.
The first task for technology implementation is to gain the trust of the patients. As for now, AI technologies are not even HIPAA compliant, so the lack of trust in technology is quite justifiable.
As artificial technology is a relatively new trend for medical imaging, there are yet no standards for its implementation. This leads to a trial-and-error method, which significantly slows the process of AI adoption. As not many companies are willing to risk their resources for this technology right away, some choose to keep their hand on the pulse but not directly involve themselves yet.
Best AI-based medical imaging software
Let’s take a more practical look at the already existing AI medical imaging companies and the services they offer. In this section, we will tell you about interesting solutions that are changing medical imaging right now.
Volpara Health is an AI-based medical imaging solution for early diagnosis of breast cancer. More than 16.5 million patients have assessed their breast density and fibroglandular tissue thanks to Volpara. Powered by machine learning, Volpara provides their customers with 4 useful algorithms:
- TruDensity. Based on the provided mammogram scans, Volpara assesses the volumetric breast density (VBD%) in a typical range from 2 to 35%. This precise number helps radiologists to process the sensitivity of the mammogram and optimize the screening, considering the breast density and its difference.
- TruPGMI. This is an image evaluation algorithm that makes sure of the correct positioning during the imaging so that all of the breast tissue is scanned. This algorithm categorizes positioning as Perfect (P), Good (G), Moderate (M), or Inadequate (I) to create an overall positioning assessment.
- TruRadDose. Instead of sticking to the equipment manufacturer's recommendations, it helps analyze the required dose of radiation based on the patient’s breast density. Thus, it helps avoid excessive radiation use in cases when a lower dosage is applicable effectively.
- TruPressure. This algorithm measures the compression pressure that is applied to the breast to help technologists form a better understanding of the patient experience and estimate the effectiveness of particular mammograms.
With a patient-oriented approach, Volpara became a popular software for medical imaging with more than 400 peer-reviewed publications mentioning the solution.
Qure AI is an example of AI with a deep learning algorithm that helps clinicians with diagnoses and patients with personalized treatment. Helping more than 10 million people around 70 countries, the solution offers various AI-based services with algorithms trained on 8+ million scans. Let’s take a closer look at each of them:
- qXR. AI for chest X-rays. A solution trained on over 4.2 million X-rays helps accurately find and classify lung, bones, heart, and diaphragm lesions in less than a minute. Such quick responses derive from the algorithm that initially separates normal X-rays from suspicious ones not to waste time reporting healthy scans.
- qER. AI for head CTs. qER is able to identify 11 various critical cases on non-contrast head CT and immediately alert the neurocritical care department about it. The tool recognizes fractures, infarcts, and abnormalities, highlights their spots on an overlay, and delivers them to specialists for timely care.
- qCT. AI for lung nodules. The qCT algorithm trained on more than 200,000 chest CTs can identify, monitor and characterize lung nodules found on provided scans. Along with being capable of distinguishing lung lesions from complex anatomical structures on lung CTs, this solution helps clinicians identify lung cancer at its early stages.
- qVH. AI for the ultrasound. This algorithm provides patients with enhanced care for cardiovascular diseases. Automated detection of carotid artery stenosis, and atherosclerotic plaques along with analyzing the potential risks for asymptomatic patients makes it efficient in terms of timely care and early discovery.
- qMSK. AI for musculoskeletal X-rays. The qMSK algorithm is similar in pattern recognition to the qXR one, but the area of specialization differs significantly. This solution is also capable of detecting and classifying issues in less than a minute, however, it aims at a greater area than the chest only. Bone fractures and joint dislocations across the whole skeleton are now easily detected and identified.
Armed with machine learning, deep learning, and computer vision, AI technologies are knocking on the door to aid radiologists and save time for their patients. AI has already found its applications in the early detection and classification of brain tumors, breast cancer, neurological diseases, cardiovascular issues, and musculoskeletal injuries based on the provided MRI, CT, and X-ray scans.
Despite being quite challenging, there are significant benefits to adopting AI for medical imaging. Quicker diagnosis, enhanced analysis, and deeper lesion classification are heavy arguments for implementing artificial intelligence into your business.
We at Mind Studios enjoy helping our customers with unique healthcare solutions and will be happy to help you pick the right AI solution for your exact business needs. In case you have any questions or a great idea in mind – feel free to contact us, and our analysts will reach out to you shortly.