Summary:
UCLA researchers from the Departments of Anesthesiology & Perioperative Medicine, Ophthalmology, Computational Medicine, and Computer Science have developed a novel uniform framework to diagnose disease-related risk factors in any volumetric medical imaging data.
Background:
Automated diagnosis of medical imaging is undoubtedly important as in most cases, in addition to reducing costs and treatment burden, it may ameliorate patient care. Many medical imaging diagnoses rely on 3D medical imaging, such as magnetic resonance imaging (MRI) scans, optical coherence tomography (OCT) scans, and ultrasound (US) videos. Training an AI model to accurately measure disease-related risk factors in such volumetric medical imaging generally requires a large annotated dataset. However, the procurement of annotated volumetric-medical-imaging datasets is expert-time prohibitive (far more than 2D medical imaging) and thus, the existence of such datasets is very limited. Consequently, even if a transfer learning approach is being taken, conventional 3D-based AI vision models are practically bounded by a performance ceiling. Several attempts were undertaken to tackle volumetric medical imaging tasks with sparsely annotated training datasets on different data modalities. The main limitation of these approaches is that due to the limited data, they are tailored and optimized for a specific medical data modality. While each modality requires a specific treatment, there are commonalities across the different data modalities, and leveraging them can provide improved results across multiple modalities (even when data are limited) and a faster development time for future predictive models.
Innovation:
Dr. Avram and his research team have developed an innovative framework called SLice Integration by Vision Transformers (SLIViT) for detecting clinical features in volumetric imaging data. SLIViT employs cutting-edge computer vision techniques to extract features from each 2D layer and comprehensively integrates them to generate a single diagnostic prediction. Importantly, the unique architecture of the model allows pre-training on 2D annotated medical imaging data, which are more affordable to annotate and thus more accessible. The pre-training enables SLIViT to effectively learn visual features and leverage them to make predictions on 3D imaging data, which typically has fewer available annotations. SLIViT surpasses domain-specific state-of-the-art computer vision models in various learning tasks and data modalities demonstrating its utility and generalizability. When trained on less than 700 annotated volumes, SLIViT was also able to reach trained clinical specialists' performance, 5,000x faster illustrating its potential to reduce the burden on clinicians and expedite ongoing research.
Press Release:
UCLA Health: New AI model efficiently reaches clinical-expert-level accuracy in complex medical scans
Publication:
Accurate prediction of disease-risk factors from volumetric medical scans by a deep vision model pre-trained with 2D scans
Demonstration Video:
An AI method for diagnosing disease risk factors in 3D medical imaging data presented @ AI Week 2023
Potential Applications:
- Volumetric imaging diagnostics (MRI, OCT, ultrasound)
- Disease quantification (cancer, retinal, cardiac, hepatic, and others)
- Treatment strategy improvement
- Automated patients’ prioritization
Advantages:
- Leverages existing (and more affordable) 2D annotated data
- Outperforms existing domain-specific computer vision models
- Compatible with any volumetric data modality and learning task
- Robust to within-volume frames permutation (allowing it to handle chaotic datasets)
Development to Date:
A series of benchmarking comparisons were conducted to evaluate the performance of SLIViT in comparison to other models (SLIVER-net, EchoNet, and 3D ResNet) across various classification and regression tasks of retinal disease, cardiac function diagnosis, and hepatic fat content measurement using only the hundreds to thousands available annotated volumes per learning task. The results consistently showed that SLIViT outperformed all the other domain-specific state-of-the-art models.
Reference:
UCLA Case No. 2023-155