Deep learning-based correction of cataract-induced influence on macular pigment optical density measurement by autofluorescence spectroscopy

**Background:** Macular pigment (MP), composed of lutein, zeaxanthin, and meso-zeaxanthin, absorbs blue light and quenches oxygen radicals, protecting the retina from photooxidative damage. Accurate measurement of macular pigment optical density (MPOD) is important for research and clinical studies on age-related macular degeneration. Autofluorescence spectroscopy using SPECTRALIS is a common objective method, but cataracts cause underestimation of MPOD due to absorption and scattering of blue excitation light by the opacified lens. Previous work proposed a deep learning (DL) correction method validated by internal cross-validation; this study aims to validate it using an external dataset.

**Methods:** This cross-sectional study included 157 eyes of 105 patients (mean age 73.0 ± 9.0 years, 71 male/86 female eyes) who underwent cataract surgery with yellow-tinted intraocular lens implantation. MPOD at 0.25°, 0.5°, 1°, and 2° eccentricities and macular pigment optical volume (MPOV) within 9° eccentricity were measured using SPECTRALIS before and 4-5 days after surgery. Autofluorescence images were subjectively classified as high (14%), moderate (26%), or poor (60%) quality. A DL model (pre-trained VGG16 fine-tuned) was trained on a separate dataset of 197 eyes to predict correction factors (CFs) from pre-operative images. The model was then applied to the validation dataset. Errors were calculated as (corrected value – post-operative value)/post-operative value. Statistical comparisons used linear mixed models with Tukey's test.

**Key Results:** Pre-operative MPOD and MPOV were significantly lower than post-operative values (P < 0.001 for all). Mean uncorrected errors ranged from 21.2% (MPOV) to 39.0% (MPOD at 0.25°). After DL correction, mean absolute errors were: MPOD at 0.25°: 12.4 ± 10.9% (median 9.5%), at 0.5°: 11.5 ± 10.9% (median 8.9%), at 1°: 11.7 ± 14.6% (median 8.0%), at 2°: 9.3 ± 8.5% (median 6.6%), and MPOV: 8.2 ± 8.2% (median 5.8%). All corrected errors were significantly lower than uncorrected errors (P < 0.001). For high and moderate quality images, mean errors were significantly lower than for poor quality images for MPOD at 0.25° (8.9% vs 14.7%, P=0.002), 0.5° (8.5% vs 13.5%, P=0.009), 2° (7.6% vs 10.5%, P=0.033), and MPOV (6.0% vs 9.7%, P=0.0054), but not for MPOD at 1° (11.4% vs 11.8%, P=0.50). Predicted CFs increased toward the foveal center (mean CF at 0.25°: 1.67, at 2°: 1.32).

**Clinical Implications:** The DL correction method effectively reduces cataract-induced underestimation of MPOD and MPOV, with errors comparable to those from internal cross-validation. It is easy to apply and particularly suitable for estimating MPOV in eyes with relatively good autofluorescence image quality (high or moderate). However, correction is less effective for poor-quality images and for MPOD at eccentricities close to the foveal center. The authors propose a hybrid approach combining subjective image quality classification with DL correction. Limitations include single-center design, Japanese population, and potential influence of surgical inflammation. Future studies should include multi-facility, multi-racial datasets and explore self-supervised learning using post-operative images.