Deep learning predicts disease risk from 3D medical images
The SLice Integration by Vision Transformer model accurately predicts disease risk factors by analyzing scans across various volumetric medical imaging modalities.
Researchers from the University of California, Los Angeles have developed an AI tool that uses 3D biomedical imaging data to predict disease risk factors, according to a recent study in Nature Biomedical Engineering.
The UCLA research team designed the model, called SLice Integration by Vision Transformer (SLIViT), to use artificial neural networks in order to address a common hurdle to the development of AI models for 3D imaging analysis: a lack of large data sets.
Standard 2D medical images show length and width, but 3D images add depth, allowing for further insights into the organ or disease captured by the scan. However, these volumetric images take more skill, attention and time for expert clinicians to interpret -- the researchers indicated that a 3D retinal scan, for example, might be composed of one hundred 2D images.
This imaging modality would require close inspection by an expert to find potentially subtle disease biomarkers, such as the volume of swelling in the body part being scanned. Doing this at a large enough scale to create robust 3D medical imaging data sets for AI training presents a challenge for researchers.
"While there are many AI methods for analyzing 2D biomedical imaging data, compiling and annotating large volumetric datasets that would be required for standard 3D models to exhaust AI's full potential is infeasible with standard resources. Several models exist, but their training efforts typically focus on a single imaging modality and a specific organ or disease," explained study co-first author Oren Avram, Ph.D., a postdoctoral researcher at UCLA Computational Medicine, in a press release.
Research team says SLIViT matched the accuracy of clinical experts when analyzing optical coherence tomography scans, ultrasound videos of the heart, 3D magnetic resonance imaging scans of the liver and 3D chest computed tomography scans.
SLIViT uses deep learning and pretraining on 2D images to overcome this.
"SLIViT overcomes the training dataset size bottleneck by leveraging prior 'medical knowledge' from the more accessible 2D domain," said co-first author Berkin Durmus, a Ph.D. student affiliated with the UCLA Samueli School of Engineering.
The model utilizes that prior knowledge to predict disease factors using labeled, moderately sized data sets. By preprocessing these 3D scans into 2D images and extracting relevant image features, the tool can also generalize across different medical imaging techniques.
When tested, SLIViT matched the accuracy of clinical experts when analyzing optical coherence tomography scans, ultrasound videos of the heart, 3D magnetic resonance imaging scans of the liver and 3D chest computed tomography scans.
"We show that SLIViT, despite being a generic model, consistently achieves significantly better performance compared to domain-specific state-of-the-art models. It has clinical applicability potential, matching the accuracy of manual expertise of clinical specialists while reducing time by a factor of 5,000. And unlike other methods, SLIViT is flexible and robust enough to work with clinical datasets that are not always in perfect order," Durmus continued.
The research team further emphasized that the tool's high performance under real-world conditions, using less data than comparable AI models, gives it a significant potential advantage over traditional 3D biomedical imaging annotation approaches.
The researchers also noted that SLIViT's efficiency could improve diagnostic timeliness, reduce data acquisition costs and expedite medical research.
"When a new disease-related risk factor is identified, it can take months to train specialists to accurately annotate the new factor at scale in biomedical images," stated Eran Halperin, Ph.D., professor of computer science and computational medicine at UCLA "But with a relatively small dataset, which a single trained clinician can annotate in just a few days, SLIViT can dramatically expedite the annotation process for many other non-annotated volumes, achieving performance levels comparable to clinical specialists."
Shania Kennedy has been covering news related to health IT and analytics since 2022.
