A deep learning system using ultra-widefield fundus imaging exhibited relatively good performance in detecting 5 common retinal lesions in rural areas.
A recent diagnostic study reported the positive performance of an ultra-widefield (UWF) fundus image-based deep learning system as a screening tool for 5 retinal lesions in a rural setting, but noted decreases compared with the model development stage.1
Investigators from Sun Yat-Sen University suggested the differences in image quality, lesion proportion, and complexity of lesion composition between model development and the rural screening stage were factors involved with the model’s performance.
“The deep learning system exhibited relatively good performance for detecting 5 common retinal lesions,” wrote the investigative team, led by Haotian Lin, MD, PhD, and Xioaling Liang, MD, PhD, Zhongshan Ophthalmic Center, Sun Yat-sen University. “However, the model performance decreased to some extent compared with that in the model development stage.”
Rural populations can be particularly vulnerable to retinal diseases, given their relatively low socioeconomic status, limited medical knowledge, and low awareness of eye care.2 Accessibility to timely screening and treatment is also low in rural areas, given the shortage of experienced professionals, meaning those in the greatest need were less likely to undergo screening.
Artificial intelligence, including deep learning systems, has grown in interest as a means to address treatment accessibility challenges. Deep learning systems have shown high accuracy in screening for retinal diseases, but most models for retinal diseases were developed using fundus images with a 45° to 55° field of view, leaving the peripheral retina unscreened.3
UWF fundus imaging has a 200° field of view, covering nearly 80% of the retinal area in a single capture without pupillary dilation. Lin and colleagues previously developed a deep learning system that can accurately detect 5 common retinal lesions from UWF fundus images – however, they noted its effectiveness for population screening in rural areas warrants further explanation before large-scale implementation.4
The prospective, rural screening diagnostic study was conducted between November 2020 - March 2021 in 24 villages in Yangxi, Guangdong Province, China.1 The previously developed system based on UWF fundus images screened for 5 retinal lesions, including retinal exudates or drusen, glaucomatous optic neuropathy, retinal hemorrhage, lattice degeneration or retinal breaks, and retinal detachment.
Model performance in rural screening was tested using prospectively collected UWF fundus images and then compared to the previous model development stage, to elucidate model performance improvements. The analysis also compared image quality, lesion proportion, and complexity of lesion composition between the rural screening and model development stages.
Captured images were analyzed by both the deep learning system and trained ophthalmologists. The presence or absence of the 5 relevant types of retinal lesions in the primary artificial intelligence report was used to validate the deep learning system in rural screening, including its accuracy, sensitivity, specificity, receiver operating characteristic (ROC) curve, and area under the receiver operating characteristic curve (AUC).
Overall, 6222 UWF images from 6222 eyes in 3149 participants were screened, including 1685 women (53.5%), with a mean age of 70.9 years. Retinal exudates or drusen had the highest proportion among the 5 retinal lesions (46.4%).
Upon analysis, the DLS achieved a mean AUC of 0.918 (95% CI, 0.892 - 0.944) for detecting 5 retinal lesions in the entire data set when applied to patients in rural areas. The mean AUC in the rural stage was lower than in the model development stage (0.998; 95% CI, 0.995 - 1.000; P <.001).
In comparison to images in the model development stage, the fundus images in the rural screening had an increased frequency of poor quality (13.8% [n = 860] vs. 0%), increased variations in lesion proportions (0.1% [n = 6 of 6222] to 36.5% [n = 2271 of 6222] vs. 14.0% [n = 2793 of 19,891] to 21.3% [n = 3433 of 16,138), and increased complexity of lesion composition.
Compared with the fundus images in the model development stage, the fundus images in this rural screening study had an increased frequency of poor quality (13.8% [860 of 6222] vs 0%) and increased variation in lesion proportions (0.1% [6 of 6222]-36.5% [2271 of 6222] vs 14.0% [2793 of 19 891]-21.3% [3433 of 16 138]).
The investigative team found the fundus images in the rural screening had an increasing complexity of lesion composition. The mean detection accuracy for the 5 lesions was found to decrease gradually, from 0.934 to 0.743 as the number of lesions in a single image increased.
“Complex disease features may affect each other, and relatively minor pathologic changes can be obscured by obvious features, leading to a certain degree of misdiagnosis and missed diagnosis,” they wrote. “Further study is needed to interpret the association between the complexity of lesion composition and model performance.”