New AI Framework Achieves Breakthrough in 3D Medical Image Analysis

Researchers at UCLA have developed an AI system that quickly analyses 3D medical images with expert-level accuracy, promising faster diagnostics.

Researchers at UCLA have developed an advanced deep-learning framework that can swiftly teach itself to analyse and diagnose 3D medical images such as MRIs with an accuracy that rivals expert clinicians. This new system, described in their paper “Accurate prediction of disease-risk factors from volumetric medical scans by a deep vision model pre-trained with 2D scans” published in Nature Biomedical Engineering, promises to enhance diagnostic efficiency and speed, potentially transforming clinical workflows.

Unlike other AI models tailored for specific imaging types, the UCLA framework exhibits broad adaptability across various imaging modalities. It has been tested on 3D retinal scans, ultrasound videos, MRI scans for liver disease, and CT scans for detecting chest nodule malignancy. This versatility sets it apart, offering a valuable tool for many clinical settings. Ongoing studies are planned to explore further applications of this technology.

Artificial neural networks, like the one used in the UCLA model, train themselves through repeated calculations and the analysis of large datasets labelled by clinical specialists. While standard 2D imaging, which shows only length and width, has been extensively studied, 3D images present an additional challenge due to their depth. For instance, a single 3D retinal scan can comprise almost 100 2D images, demanding meticulous review by skilled professionals to detect subtle disease biomarkers.

Dr Oren Avram, a postdoctoral researcher at UCLA Computational Medicine and co-first author of the study, said, “Several models exist, but their training efforts typically focus on a single imaging modality and a specific organ or disease.” Avram emphasised that the new framework overcomes these limitations, offering broader applicability.

How SLIViT Works

The UCLA team’s model, called SLIViT (SLice Integration by Vision Transformer), uses a combination of two artificial intelligence components and a novel learning method. These features enable it to predict disease risk factors from medical scans across different 3D imaging types while requiring relatively smaller labelled datasets than conventional models.

“SLIViT overcomes the training dataset size bottleneck by leveraging prior ‘medical knowledge’ from the more accessible 2D domain,” said Berkin Durmus, a UCLA PhD student and co-first author of the article. Durmus also highlighted that SLIViT “consistently achieves significantly better performance compared to domain-specific state-of-the-art models,” reducing time while maintaining accuracy.

One of the framework’s strengths is its flexibility to work with clinical datasets that are not perfectly structured. According to Durmus, this gives it a practical edge in real-world scenarios, where data imperfections are common.

Potential Clinical Impact

The implications of this technology are profound. “SLIViT thrives with just hundreds – not thousands – of training samples for some tasks, giving it a substantial advantage over other standard 3D-based methods in almost every practical case related to 3D biomedical imaging annotation,” said Dr SriniVas R. Sadda, professor of Ophthalmology at UCLA and the Artificial Intelligence & Imaging Research director at the Doheny Eye Institute. This could expedite diagnoses and reduce the burden on healthcare professionals.

Eran Halperin, professor at UCLA’s Henry Samueli School of Engineering, added that “even if financial resources were unlimited, ongoing research will always face challenges posed by limited training datasets.” He explained that SLIViT’s ability to work with smaller datasets could “dramatically expedite the annotation process” and reduce the time required to achieve high levels of accuracy in clinical settings.

Moving forward, the UCLA researchers plan to explore SLIViT’s use in predictive disease forecasting, potentially improving early diagnosis and treatment planning. They also aim to ensure that systematic biases in AI models do not exacerbate healthcare disparities.

Future Applications for Complex Data

While SLIViT has already demonstrated impressive performance with 3D medical imaging, its success prompts a question: could this model be adapted to handle other complex scientific datasets beyond healthcare? For instance, fields like genomics, climate modelling, and particle physics generate vast volumes of intricate data that often overwhelm traditional analysis methods. If SLIViT can process medical imaging so effectively, might it also offer new possibilities for managing and interpreting diverse forms of scientific research data? Exploring these avenues could open new frontiers for AI in scientific discovery.