Artificial intelligence has been revolutionizing the medical domain. It is the branch of computer science dealing with intelligent computing.
Researchers began taking an interest in artificial intelligence (AI) for life sciences during the 70s. Earlier works focused more on chemistry than medicine, such as the Dendral project.
Modern artificial intelligence aims to solve practical healthcare problems. Advanced techniques like deep learning have maximized the impact of technology in healthcare, particularly the use of AI in radiology.
Radiology is the field of medical science that uses radiation to generate medical imaging, e.g., X-ray, CT scans, ultrasound, and MRI images, to detect deformities and tumors. AI algorithms can automatically detect complex anomalous patterns in image data to provide an assistive diagnosis for patients.
American Department of Radiology suggests AI adoption in radiology went up from zero to 30% from 2015 to 2020. A slow but steady growth.
In this article, we’ll take a closer look at the impact of AI on the field of radiology.
Here’s what we’ll cover:
Solve any video or image labeling task 10x faster and with 10x less manual work.
Don't start empty-handed. Explore our repository of 500+ open datasets and test-drive V7's tools.
Ready to streamline AI product deployment right away? Check out:
So, what is AI in Radiology?
Radiology generates a lot of clinical image data. Radiologists must spend hours sorting through these images, writing analyses, and finalizing the diagnosis.
But medical images can also be processed and analyzed by computer vision (CV)–a specialized field of artificial intelligence to predict diseases accurately.
AI brings several benefits for Radiologists, which ease their work. Some key benefits include:
Today, deep learning-based specialized CV algorithms have reached such intelligence that they can distinguish the slightest anomalies and produce accurate classifications comparable to that of humans, at times even better.
Other than classification, deep learning architectures such as the U-Net specialize in the automated segmentation of medical images. Segmentation enhances image analysis and aids practicing radiologists. These models offer radiologists a second opinion regarding analysis and add confidence to their diagnosis. They may even point out anomalies that may not be obvious to the naked eye.
3D modeling also benefits from artificial intelligence. Models can segment medical images with precision and fuse multiple segments, fed to 3d rendering software for reproduction. Radiologists can study these models for additional analysis.
Not only are they accurate but AI models perform these tasks, with the correct hardware, within seconds. This speeds up radiology practice and reduces stress on the practitioner.
However, not all that glitters is gold. Artificial intelligence methods benefit radiology practice, but there are also certain concerns.
Martin-valdivia and Luna present a SWOT analysis for AI in medical imaging. They highlight the lack of standardization as a potential weakness of AI in radiology. The lack of standardized benchmarks makes it difficult to compare or authenticate the performance of any model. Without validation, it is difficult to decide whether a model is ready for practical implementation.
The researchers further comment on model explainability as a threat to artificial intelligence in medical science. An AI algorithm's interpretability is critical for clinical data science. Deep learning algorithms implement a neural network architecture and process data sets through several thousand neurons. It is impossible for humans to derive concrete logic behind such complex mathematics. The lack of reasoning raises questions about the reliability of AI models. Artificial intelligence interpretability becomes more crucial in clinical integration because the slightest mistake can have adverse consequences. For people to trust computer-aided detection and decisions, some form of reasoning is required.
The validation of AI in radiology depends heavily on the availability of patient data. Generating validation data sets is a time-consuming task that bottlenecks many machine learning projects. However, AI models can not be tested for modern applications without such data sets.
The fact that medical researchers have access to patients' personal records for training models stirs up controversy. This does not fare well with people who value their privacy and acts as a hurdle for the practical implementation of AI.
The correct use of data, data privacy, and data biases all fall under the umbrella of the ethics of AI. Ethics are vital to ensure that no harm comes from practicing AI. Let's explore this further in the next section.
Ethics plays an important part in artificial intelligence, and the importance jumps up when we factor in the field of medicine. The whole purpose of AI is to make decisions without human intervention. It applies complex logic, but that is where the problem starts.
Basing medical decisions on mathematics from an AI algorithm is not something every patient will appreciate.
The European Society of Radiology reports, “We must ensure that radiological AI remains human-centric, helps patients, contributes to the common good, and evenly distributes both the benefits and harms that may occur.”
As practitioners, radiologists have access to patients' records. This data can be easily misused in the form of identity theft, allowing the thief to carry out insurance fraud and illegally obtain prescription drugs. As Brady and Neri (2020) highlight in their work, radiologists are responsible for ensuring that patient data is used only for the good of patients and clinical practice in general.
Responsible use of radiology data sets also means taking care of any biases. Biases harm the patient population based on ethnicity, gender, or cast. While preparing a medical dataset, it may not be possible to include patients from all ethnicities. When models trained on such data are implemented, they provide inaccurate results for whatever population was missing in the training data set. The responsibility for unbiased data falls to data scientists who use it for machine learning applications.
Geis, et al. (2019) also talk about the ethic of practice in artificial intelligence and radiology workflow. They talk about the error in validating machine learning algorithms via omission and commission. Omission occurs when radiologists fail to recognize erroneous outputs from AI software. Commission occurs when one willingly accepts the decision of an AI software despite other contentions.
This is called automation bias and is more prevalent in resource-poor countries where a sufficient number of radiologists may not be available.
Ethics of artificial intelligence also include the interpretability of AI software.
Understanding how AI systems reach conclusions would help track biases within the calculation. Brady and Neri (2020) pose an important question, "Will the public accept imperfections in AI-driven healthcare as it relates to individuals in favor of a potential greater good for the population at large?"
Though such questions have no easy answer, there have been efforts to improve model explainability, especially for convolutional neural networks.
Source: Research by Zeiler and Fergus (2013) demonstrates the features that a CNN extracts from image data
AI in Radiology brings a particular set of responsibilities. It requires certain standards and rules to be followed. Regulations to ensure that no harm may come out of it. While on the topic of uses, let’s see how AI helps radiologists solve problems.
We have discussed the challenges, but how is AI used in radiology?
Perhaps it’s time we looked at some practical applications.
Neurological disorders occur when certain brain parts stop functioning correctly. These parts are primarily responsible for memory and speech and cause diseases like Alzheimer's and Parkinson’s.
Using convolutional neural networks, artificial intelligence can extract meaningful information from brain image data. This information helps detect irregular brain development. Research at Mount Sinai Health Systems demonstrated that novel artificial intelligence techniques could be used to identify the causes of Alzheimer's. They analyzed human brain images using deep learning techniques and monitored staining intensity.
Moreover, researchers can also diagnose these diseases by tracking the patients' retinal movement.
The conventional method of dealing with brain tumors is very time-consuming.
Classifying the tumor takes up to 40 mins, and only after that can doctors decide on a course of action for further treatment.
Brain tumors can be identified and classified using MRI images and machine learning. The results are achieved within minutes and with high accuracy. Precise data annotation is required to achieve accurate results. V7labs auto-annotation feature not only saves time but yields amazing results.
A recent study used CNNs with MRI scans as the input image and achieved an accuracy of 98.56% in classifying brain tumor types. Another AI research in the UK discovered a non-invasive technique for classifying tumors in children. They used the diffusion of water molecules to obtain contrast in MRI scans. Later, the apparent diffusion map is extracted and fed to an AI model.
Breast cancer detection occurs via a careful examination of the mammography report, MRI scan, or Ultrasound. The manual verification process accompanies the risk of human error, and misdiagnosis is quite common.
AI tools can enhance mammography examinations.
AI tools can enhance mammography examinations. A study published in the Radiology Society of North America (RSNA) aimed to detect breast cancer risk by examining mammograms. 87.6% of the screen-detected cancers scored the highest risk.
They concluded the research by adding, “In our study, we assumed that all cancer cases selected by the AI system were detected. This might not be true in a real screening setting. However, given that assumption, AI will probably be of great value in the interpretation of screening mammograms in the future.”
This use case differs from the others. Radiology requires the patient to be exposed to harmful radiation to get a decent MRI or CT scan. Radiologists cannot discern any anomaly from normal tissues without proper medical imaging. The longer the exposure, the better the image quality. Adults may not be concerned, but it can be harmful to children.
Using upscaling models, artificial intelligence can enhance the resolution of these images. Data for such models can be synthetically generated via image data augmentation techniques. These enhanced images can be fed to other AI tools for carrying out the radiology workflow.
The past and present advancements in artificial intelligence for healthcare are evidence enough to believe that the future is promising.
Savadjiev, et al. (2018) analyzed various machine and deep learning algorithms used for clinical research. From the symbolic interpretation of images to deep learning, they analyzed the strength and limitations of each algorithm.
They concluded that the question is not whether AI in radiology is mature enough but which clinical tasks in radiology are the most and least likely to benefit from AI.
Radiologists are adopting AI tools, but there is still a long way to go. Charlene Liew (2018) discusses a strategy for merging AI with radiology workflows. They highlight a few routine tasks that can be automated, including
These use cases have already seen the light of day and have achieved promising accuracies.
They also ponder the possibility of unemployment due to artificial intelligence gaining traction. Although it is a valid concern, many argue that artificial intelligence is only meant to support radiologists, not replace them.
The only worry for practitioners is that radiologists who use AI will certainly replace those who don’t.
AI is here to stay, and the sooner radiologists start benefiting from it, the better.
Artificial intelligence techniques for medical imaging yield promising results. These techniques hold the potential to assist radiologists in their day-to-day tasks by processing quicker results and offering a second opinion.
Some typical applications include the classification of tumors and other irregularities in the brain. It can also be used for optimizing radiation dosage.
Automating medical decisions brings a particular set of responsibilities. Any data provided to engineers is private and must only be used for the benefit of medical science. AI must also be wary of model interpretability to establish mutual trust.
Explore other articles:
“Collecting user feedback and using human-in-the-loop methods for quality control are crucial for improving Al models over time and ensuring their reliability and safety. Capturing data on the inputs, outputs, user actions, and corrections can help filter and refine the dataset for fine-tuning and developing secure ML solutions.”
Building AI products? This guide breaks down the A to Z of delivering an AI success story.