Scholarlink

Aims <p>To develop and validate a low-cost, portable automated refraction system (Tongren Digital Sight, TRDS) using deep learning and infrared eccentric photorefraction to improve refractive error correction in resource-constrained settings.</p> Methods <p>This randomised controlled crossover trial enrolled 282 participants (18–60 years of age) at Beijing Tongren Hospital. TRDS utilised low-cost components (complementary metal-oxide semiconductor (CMOS) detector, infrared light emitting diodes) and an end-to-end deep learning model trained on 362,000 images to analyse fundus reflection signals. Primary outcomes included agreement between TRDS and subjective refraction (gold standard) via Pearson correlation, Bland–Altman plots and intraclass correlation coefficients (ICC) for spherical equivalent (<i>M</i>), horizontal (<i>J</i><sub>0</sub>) and oblique (<i>J</i><sub>45</sub>) cylinder parameters. Secondary outcomes included corrected visual acuity.</p> Results <p>TRDS showed strong correlations with subjective refraction: <i>M</i> (<i>r</i> = 0.96, ICC = 0.96), <i>J</i><sub>0</sub> (<i>r</i> = 0.74, ICC = 0.71) and <i>J</i><sub>45</sub> (<i>r</i> = 0.76, ICC = 0.75). Mean differences were 0.07 D (<i>M</i>), −0.06 D (<i>J</i><sub>0</sub>) and 0.00 D (<i>J</i><sub>45</sub>), with 95% limits of agreement within clinical tolerance. Corrected visual acuity was non-inferior to that of subjective refraction (mean difference = 0.03 logMAR, <i>p</i> &lt; 0.01). Subgroup analysis showed consistent performance across myopia and hyperopia, except for moderate agreement in high myopia (<i>M</i>: <i>r</i> = 0.62). Subjective and TRDS refraction showed corrected visual acuity of 0.01 (SD = 0.13) and −0.02 (SD = 0.08) logMAR, respectively, with a significant mean difference of 0.03 logMAR (95% LoA: 0.01–0.04, <i>p</i> &lt; 0.01). Corrected visual acuity &gt;0.18 logMAR occurred in 13 (4.6%) and 14 (5.0%) participants for subjective and TRDS refraction, respectively.</p> Conclusions <p>The deep learning-empowered TRDS system demonstrated high accuracy and portability. It offers a scalable solution to improve refractive error correction rates, aligning with WHO goals to enhance global eye health.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning–empowered low-cost portable automated refraction system: A solution to the inadequate effective correction rate of refractive errors in resource-limited areas

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月24日

查看全文
  • Huang Yan,
  • Zhen Yi,
  • Sun Qilin,
  • Chang Hong,
  • Huang Yan,
  • Tang Wei

摘要

Aims

To develop and validate a low-cost, portable automated refraction system (Tongren Digital Sight, TRDS) using deep learning and infrared eccentric photorefraction to improve refractive error correction in resource-constrained settings.

Methods

This randomised controlled crossover trial enrolled 282 participants (18–60 years of age) at Beijing Tongren Hospital. TRDS utilised low-cost components (complementary metal-oxide semiconductor (CMOS) detector, infrared light emitting diodes) and an end-to-end deep learning model trained on 362,000 images to analyse fundus reflection signals. Primary outcomes included agreement between TRDS and subjective refraction (gold standard) via Pearson correlation, Bland–Altman plots and intraclass correlation coefficients (ICC) for spherical equivalent (M), horizontal (J0) and oblique (J45) cylinder parameters. Secondary outcomes included corrected visual acuity.

Results

TRDS showed strong correlations with subjective refraction: M (r = 0.96, ICC = 0.96), J0 (r = 0.74, ICC = 0.71) and J45 (r = 0.76, ICC = 0.75). Mean differences were 0.07 D (M), −0.06 D (J0) and 0.00 D (J45), with 95% limits of agreement within clinical tolerance. Corrected visual acuity was non-inferior to that of subjective refraction (mean difference = 0.03 logMAR, p < 0.01). Subgroup analysis showed consistent performance across myopia and hyperopia, except for moderate agreement in high myopia (M: r = 0.62). Subjective and TRDS refraction showed corrected visual acuity of 0.01 (SD = 0.13) and −0.02 (SD = 0.08) logMAR, respectively, with a significant mean difference of 0.03 logMAR (95% LoA: 0.01–0.04, p < 0.01). Corrected visual acuity >0.18 logMAR occurred in 13 (4.6%) and 14 (5.0%) participants for subjective and TRDS refraction, respectively.

Conclusions

The deep learning-empowered TRDS system demonstrated high accuracy and portability. It offers a scalable solution to improve refractive error correction rates, aligning with WHO goals to enhance global eye health.