AI detects eye disease and risk of Parkinson’s from retinal images
来源：Nature | 发布时间：2023-09-22
Scientists have developed an artificial intelligence (AI) tool capable of diagnosing and predicting the risk of developing multiple health conditions — from ocular diseases to heart failure to Parkinson’s disease — all on the basis of people’s retinal images.
AI tools have been trained to detect disease using retinal images before, but what makes the new tool — called RETFound — special is that it was developed using a method known as self-supervised learning. That means that the researchers did not have to analyse each of the 1.6 million retinal images used for training and label them as ‘normal’ or ‘not normal’, for instance. Such procedures are time-consuming and expensive, and are needed during the development of most standard machine-learning models.
Instead, the scientists used a method similar to the one used to train large language models such as ChatGPT. That AI tool harnesses myriad examples of human-generated text to learn how to predict the next word in a sentence from the context of the preceding words. In the same kind of way, RETFound uses a multitude of retinal photos to learn how to predict what missing portions of images should look like.
“Over the course of millions of images, the model somehow learns what a retina looks like and what all the features of a retina are,” says Pearse Keane, an ophthalmologist at Moorfields Eye Hospital NHS Foundation Trust in London who co-authored a paper published today in Nature describing the tool. This forms the cornerstone of the model, and classifies it as what some call a foundation model, which means that it can be adapted for many tasks.
A window into human health
A person’s retinas can offer a window into their health, because they are the only part of the human body through which the capillary network, made up of the smallest blood vessels, can be observed directly. “If you have some systemic cardiovascular disease, like hypertension, which is affecting potentially every blood vessel in your body, we can directly visualize [that] in retinal images,” Keane says.
Retinas are also an extension of the central nervous system, sharing similarities with the brain, which means that retinal images can be used to evaluate neural tissue. “The rub is that a lot of the time people don’t have the expertise to interpret these scans. This is where AI comes in,” Keane says.
Once they had pre-trained RETFound on those 1.6 million unlabelled retinal images, Keane and his colleagues could then introduce a small number of labelled images — say, 100 retinal images from people who had developed Parkinson’s and 100 from people who had not — to teach the model about specific conditions. Having learnt from all the unlabelled images what a retina should look like, Keane says, the model is able to easily learn the retinal features associated with a disease.
Using unlabelled data to initially train the model “unblocks a major bottleneck for researchers”, says Xiaoxuan Liu, a clinical researcher who studies responsible innovation in AI at the University of Birmingham, UK. Radiologist Curtis Langlotz, director of the Center for Artificial Intelligence in Medicine and Imaging at Stanford University in California, agrees. “High-quality labels for medical data are extremely expensive, so label efficiency has become the coin of the realm,” he says.
The system performed well at detecting ocular diseases such as diabetic retinopathy. On a scale where 0.5 represents a model that performs no better than a random prediction and 1 represents a perfect model that makes an accurate prediction each time, it scored between 0.822 and 0.943 for diabetic retinopathy, depending on the data set used. When predicting the risk for systemic diseases — such as heart attacks, heart failure, stroke and Parkinson’s — the overall performance was limited, but still superior to that of other AI models.
RETFound is so far one of the few successful applications of a foundation model to medical imaging, Liu says.
Expanding applications
Researchers are now looking ahead to what other types of medical imaging the techniques used to develop RETFound might be applied to. “It will be interesting to see whether these methods generalize to more complex images,” Langlotz says — for example, to magnetic resonance images or computed tomography scans, which are often three- or even four-dimensional.
The authors have made the model publicly available, and hope that groups around the world will be able to adapt and train it to work for their own patient populations and medical settings. “They could potentially take this algorithm and fine-tune it, using data from their own country to have something that’s more optimized for their use,” Keane says.
“This is tremendously exciting,” Liu says. But using RETFound as the basis for other models to detect diseases comes with a risk, she adds. That’s because any limitations embedded in the tool could leak into future models that are built from it. “It is now up to the authors of RETFound to ensure its ethical and safe usage, including transparent communication of its limitations, so that it can be a true community asset.”
AI又一飞跃！从眼睛里看出帕金森病风险，还能诊断多种眼部疾病
我们常说，眼睛是我们看向世界的窗口。但鲜为人知的是，与此同时，我们的眼睛也可以成为外界洞悉我们健康状况的窗口。
记录人体健康状态的，是位于眼球后壁部的视网膜。视网膜的特殊之处在于，它是人体中唯一可以直接观察到毛细血管网络的结构。一旦我们患上系统性的心血管疾病（例如高血压），体内的每一根血管都有可能会受到影响，与病理相关的改变因此可以体现在视网膜的毛细血管网络中。此外，作为中枢神经系统的延伸，视网膜还包含了神经组织的信息，为非入侵地观察神经系统提供了可能。
了解眼睛与身体的关联，对于应对复杂疾病以及衰老相关问题有着重要意义。但对于肉眼来说，这些错综复杂的网络如同一团乱麻，我们根本无法从中读出有效的健康信息。好在，人工智能（AI）的发展提供了机遇。
在一项发表于《自然》的最新研究中，来自伦敦大学学院和Moorfields眼科医院的科学家开发了一款名为RETFound的AI视网膜基础模型，可以基于人们的视网膜图像，诊断威胁视力的眼部疾病，并且有望用于预测心血管疾病、帕金森病等多种系统性疾病的风险。伦敦大学学院和Moorfields眼科医院的周玉昆为论文第一作者兼共同通讯作者。
在这项研究之前，已经有通过视网膜图像检测疾病的AI工具出现。但这类传统的AI模型最大的问题在于，其往往需要大量带有高质量标签的图像用于训练，标签的添加需要巨大的专业医生工作量，以及高昂的费用。
相比之下，最新的RETFound工具采用了自监督学习的策略进行开发。研究团队使用了一种类似于训练ChatGPT等大语言模型的方法，利用大量未标记的视网膜图像，学习如何预测图像缺失的部分。最新研究的两个基础模型RETFound模型分别用于彩色眼底摄影以及光学相干断层扫描。这类基础模型的特点是，经过大量未标记数据的训练后，能够通过微调以适应一系列后续的检测与预测任务，具有可适应无限应用场景的潜力。
▲图：RETFound模型训练示意图
RETFound仅仅使用低至10%的人类标签，就能够匹配其他AI系统的性能，实现了效率的巨大提升。随后的测试也验证了RETFound在糖尿病性视网膜病变和青光眼等眼部疾病诊断中的良好效果。其中，在检测糖尿病性视网膜病变时，根据使用的数据集不同，评分在0.822和0.943之间（0.5分表示不优于随机预测；1分表示每次都能作出准确判断的完美模型）。
此外，研究还展示了利用RETFound预测更多系统性疾病的潜力。作者首先用160万张未标记的视网膜图像，对RETFound进行了预训练。随后，只使用少量图像，例如100位帕金森病患者的视网膜图像与100位未患病人群的视网膜图像，模型就能学习视网膜上与帕金森病相关的特征。基于这一策略，RETFound在预测心衰、心肌梗塞、中风和帕金森病等系统性疾病时，表现出了优于其他AI模型的性能。
▲图：RETFound在检测与预测不同疾病时的定量结果
领导这项研究的Pearse Keane教授表示，该研究展示了RETFound的几种应用场景案例，但它还有可能经过进一步开发，应用于数百种我们尚未探索的，会威胁视力的眼部疾病。
目前，该模型已经开源。研究团队希望，更多研究者能够通过对RETFound的训练与微调，将其应用于更广泛的应用场景。同时，研究团队正在展望，开发RETFound的技术能否推广至更复杂的医学图像，例如磁共振图像或计算机断层扫描，从而为更多疾病提供诊断或预测的工具