The emerging application of artificial intelligence (AI) and machine learning (ML) to the field of radiology is expected to bring a myriad of benefits, including augmentation and improvements in diagnosis.
However, as happens with the application of any new technology in healthcare, this future has already hit some bumps in the road. Among AI and ML’s critical needs is access to data for training, testing, validating, verifying and regulating AI- and ML-based imaging systems and tools. A paper, published in the May 2017 issue of the Journal of Digital Imaging, stated that to put AI to use in radiology there is an “urgent need to find better ways to collect, annotate, and reuse medical imaging data.”
Healthcare generates 800 billion images annually, which means there’s a lot of data out there that could be used. However, given the difficulties of image access and anonymization, algorithms are developed from relatively small sets of data so when they come out of the lab they are more like a recent grad than an experienced clinician.
Like any medical resident, algorithms need continuous training to deliver high-quality and accurate diagnosis and decisions support. Access to data, such as edge cases, is critical to creating AI and ML systems that can deliver on their promised benefits.
Increasing AI Predictive Power Requires Image Data -- Lots of it
A recent study by a Mayo Clinic and Arizona State University research team highlights the role that image data plays in developing ML-based imaging applications. The team’s research explored using ML to assess ultrasound images of thyroid nodules and determine the need for a fine-needle aspiration (FNA) biopsy. The ultimate goal of this work was to reduce the number of FNA biopsies and their costs; only 10 percent of patients selected to have a biopsy actually have cancer.
"[T]here is a clinical need for a dramatic reduction of these fine-needle aspiration biopsies," said Zeynettin Akkus, PhD from the Mayo Clinic.
Ultrasound thyroid images include features that can differentiate benign from malignant nodules and identifying those features can reduce the number of FNA biopsies. Creating a ML algorithm to identify benign nodules could eliminate the need for nearly half of FNA biopsies, while still providing high-quality care. The team had access to 100 benign cases and 50 malignant cases to develop their algorithm. This limited training dataset required artificial augmentation to train the algorithm, which is not ideal.
Ultimately, the performance of the algorithm needs improvement with more training data. "[H]opefully we can increase the [size of the] dataset to thousands of images, and…we are going to have better prediction power," explained Akkus.
AI-Ready Data Must be Findable and Accessible
Ultimately, AI systems need to be continuously trained to increase their accuracy and ability to perform accurate diagnosis. As described in the Journal of Data Imaging, an ideal imaging data set is findable and accessible.
The key is not to change the image data but to access the data through an AI-ready platform that includes features to give image data these qualities. ResolutionMDⓇ is an image viewing platform that enabled AI systems to find and access patient data no matter where the image data is located within a hospital or health system.
Findable: Images exist in silos throughout hospitals and healthcare systems. ResolutionMD provides federated image access, enabling AI systems to find image data no matter where it resides, whether on a PACS, VNA or XDS repository.
Accessible: When integrated with an AI system, ResolutionMD not only enables AI systems to find image data but also makes that data accessible to the system by eliminating any associated patient-specific data through data anonymization and de-identification.
Learn more about solving the “data starvation” problem with the ResolutionMD AI-ready platform.