Retinal Scans to Detect Peripheral Arterial Disease

Article: Eye provides clues to insidious vascular disease
Source: University of Bonn (Germany)
Published: February 11, 2022

Atherosclerosis, chronic narrowing and hardening of blood vessels by plaques, is a major risk factor for heart attacks, strokes, and death in industrialized nations. The vessel modeling changes in atherosclerosis also affect blood flow to the extremities, such as the legs and arms, resulting in conditions such as peripheral arterial disease (PAD). Because PAD does not cause symptoms in early stages, diagnosis is often made when secondary damage has already occurred; however, the consequences of progressive circulatory reduction can be dramatic or even permanent where amputation is required. Moreover, PAD significantly increases the risk of fatal heart attack and stroke, even in early stages. Researchers in Germany investigated whether prevention of cardiovascular morbidity and mortality from PAD could be detected from noninvasive methods such as routine retinal scans. They developed a deep learning software based on a convolutional neural network (CNN), which is modeled after how neural networks of the human brain work.

Attentional maps of two true positive results (left) and two true negative results (right)
showing that the CNN mostly relied on retinal vasculature to make its prediction

Because datasets for PAD are relatively scarce, the researchers first trained the model on 83,126 sample images of another vascular disease that affects the eyes, namely, diabetic retinopathy (Kaggle challenge studies in 2015 and 2019), and further validated the model on an additional 9,237 sample set. This attention-based CNN software, called Multiple Instance Learning (MIL), was then applied to color fundus photography (CFP) of 97 people suffering from PAD and compared to those of 34 healthy control subjects. The authors comment that their software also works on higher resolution photographs (EIDON widefield TrueColor confocal fundus camera) compared to most CNNs in order to discern the details of microvasculature in retinal disease. In this exploratory study, they report an ROC AUC score (a measurement of diagnostic ability in binary classifier systems) of 0.890. Senior author of the study further states, "A good 80 percent of all affected individuals were correctly identified, if we took into account 20 percent false positives...That's amazing, because even for trained ophthalmologists, PAD can't be detected from fundus images." They also report that the algorithm pays closer attention to the optic disc and temporal arcades compared to the retinal background, and that being able to visualize these attention weights gives insights into the decision-making of deep learning networks. The researchers look to expand their data set with international colleagues and further improve the performance of their diagnostic model.

My rating of this study: ⭐⭐⭐

Mueller S, Wintergerst MWM, Falahat P, et al. "Multiple instance learning detects peripheral arterial disease from high-resolution color fundus photography." Scientific Reports. 12:1389. 26 January 2022. https://doi.org/10.1038/s41598-022-05169-z