Machine learning maps which lens layers bend when your eyes focus close up

When you shift your gaze from a distant sign to the text on your phone, the lens inside each eye reshapes itself within a fraction of a second. For centuries, that reshaping has been described mostly by numbers measured at the lens surface, such as total thickness and front-surface curvature. A new study argues that this global view misses where the action actually happens. Using long-range optical coherence tomography and machine learning–assisted image segmentation, researchers mapped accommodation layer by layer through the full thickness of the human lens and found that the deformation is concentrated in the lens nucleus and the innermost cortical layers, while the outer cortex barely moves.

The study, published in the journal Eye and Vision, enrolled sixty healthy young adults with a mean age of 26 years and a range of 18 to 50 years. Each participant had one eye imaged with a commercial long-range swept-source OCT system while viewing targets through a modified Badal optometer, which allowed the investigators to set accommodative demands of 0, 2, 4, and 6 diopters without changing the apparent size of the target. The system operated at 60,000 A-scans per second, achieved an axial imaging range of 40.2 millimeters, and delivered an axial resolution of 11 micrometers, long enough to capture the entire ocular axis in a single acquisition.

That full-eye field of view matters because conventional anterior segment OCT cannot simultaneously image the whole eye and resolve the fine internal structure of the lens. The crystalline lens is not a homogeneous blob of tissue. It has a cortex surrounding a denser nucleus, a distinction first described in the eighteenth century, and its cross-sectional images show alternating light and dark bands called optical signal discontinuity zones. Under the Oxford clinical cataract classification, the cortex is divided into four zones on each side of the nucleus, labeled C1 through C4, with C4 lying innermost, closest to the nucleus. These zones reflect the layered growth pattern laid down over a lifetime of fiber cell production.

Segmenting those subtle bands automatically has been a persistent obstacle, so the team built a deep learning pipeline based on the U-Net architecture, with a pre-trained ResNet-50 encoder, a multiscale feature refinement module, and an attention-based feature integration module. Evaluated on an unseen test set of 536 images, the model reached a mean intersection over union of 0.936, with recall of 0.974 and specificity of 0.995, and per-class scores of 0.912 for the cornea, 0.913 for the lens, and 0.928 for the nucleus. The predicted surfaces were then fed into a ray-tracing algorithm that corrected the OCT images for light refraction through the cornea, aqueous humor, cortex, nucleus, and vitreous humor, using wavelength-dependent refractive indices from a published dispersion model.

Thickness measurements came from the central A-scan of each image, where a densitometric profile revealed the peaks and troughs marking the boundaries of the OSD zones, which were manually delineated and verified against the corresponding B-scans. Curvature was quantified by fitting a sphere to each refraction-corrected surface over the central 3-millimeter zone. The participants were imaged without dilating drops, and each acquisition captured two perpendicular meridians in roughly 80 milliseconds, fast enough to minimize distortion from microsaccades. Measurements at each demand were averaged across meridians and across accommodation and disaccommodation sequences, and changes were compared against the relaxed 0-diopter baseline with Bonferroni-corrected statistics.

The global biometric results matched what decades of prior work would predict. Aqueous depth and anterior chamber depth decreased progressively with demand, with effect sizes between −1.38 and −0.99, and vitreous chamber depth shortened as well, with effect sizes between −0.85 and −0.76. Per diopter of accommodation, aqueous depth changed by −16.58 micrometers and vitreous chamber depth by −6.46 micrometers. Axial length increased slightly, by about 1.12 micrometers per diopter, reaching statistical significance at 2 and 4 diopters but not at 6 diopters after correction, a borderline effect the authors interpret as a transient mechanical deformation of the posterior pole rather than true growth of the eye. Central corneal thickness showed no significant change at any demand, and pupil diameter constricted steadily, by −81.66 micrometers per diopter, the largest relative change of any biometric parameter at roughly 7.3 percent at 6 diopters.

The lens itself told a sharper story. Total lens thickness increased almost linearly at about 24 micrometers per diopter, but nuclear thickness grew even faster, at about 36 micrometers per diopter, while combined cortical thickness shrank at −14 micrometers per diopter. At the maximum 6-diopter demand, the whole lens had thickened by up to 3.7 percent, the nucleus by up to 10.1 percent, and the cortex had thinned by 5.5 percent. In other words, the lens does not swell uniformly when it focuses. The dense central mass expands and pushes outward while the surrounding cortical shell is squeezed thinner, a pattern consistent with mechanical models in which the capsule and peripheral cortex mainly transmit zonular forces and impose boundary conditions on a deforming central mass.

The zonal analysis went further and located the deformation with surprising precision. Across most of the anterior C1 through C3 zones and the posterior C2 and C3 zones, mean thickness changes were 5 micrometers or less, statistically non-significant, and associated with negligible effect sizes. The posterior C1 zone flickered with small, isolated changes at 4 and 6 diopters, which the authors describe as weak secondary modulation. But the innermost C4 zones on both sides of the nucleus thinned monotonically and significantly, the anterior C4 at −8 micrometers per diopter and the posterior C4 at −5.52 micrometers per diopter, both with P values below 0.001. Expressed as percentages, the peripheral zones moved less than 3 percent at 6 diopters, while the anterior and posterior C4 layers thinned by roughly 7 to 8 percent at 2 diopters and by 22 to 25 percent at 6 diopters. The inner cortex, it turns out, is not a passive bystander; it deforms in coordination with the nucleus it borders.

Curvature measurements reinforced the same central-lens picture. The anterior lens, anterior nuclear, and posterior nuclear radii all decreased steeply with demand, at approximately −289, −123, and −75 micrometers per diopter respectively, all highly significant, while the posterior lens radius showed only a borderline effect of −44 micrometers per diopter that did not survive correction at higher demands. The corneal radii stayed essentially flat, confirming that the cornea is largely decoupled from accommodative effort in healthy eyes. Comparisons between horizontal and vertical meridians found no coherent orientation-related differences, supporting the common modeling assumption that the accommodating lens is approximately rotationally symmetric. Sex comparisons revealed that males had longer vitreous chambers and axial lengths as stable anatomical offsets, while females showed a thicker anterior C4 zone, slightly smaller nuclear radii, and a modestly higher rate of axial elongation per diopter, subtle dimorphisms the authors say have not been reported before and warrant further study.

The authors are careful about the limits of their data. The cohort consisted of healthy young adults with low refractive error, so the findings may not extend to highly myopic, hyperopic, or astigmatic eyes, in which accommodative dynamics can differ. The curvature values are best-fit spheres over the central 3-millimeter zone and may miss aspheric shape changes, and posterior lens curvature estimates are more sensitive to assumptions about the lens gradient refractive index than anterior ones. The refraction correction also used theoretical phase refractive indices, which the authors calculate overestimates thicknesses by roughly 0.65 to 0.72 percent uniformly, a scaling that does not affect the relative changes they report. Even so, the central message stands: accommodation cannot be fully described by global lens metrics alone. The quantitative baseline they established, and the open dataset accompanying the paper, give biomechanical and optical modelers a layered target to aim at, one that treats the lens as what it actually is, a stratified, differentially deformable structure rather than a single elastic lump.

Subject of Research: Layer-specific imaging of how the crystalline lens deforms during accommodation in healthy young adults

Article Title: Layer-specific quantitative analysis of accommodation in the human lens using long-range OCT and machine learning

Article References: Soman, K., Usman, M., Urriza-Arpal, R., Gondek, G., Karnowski, K., Ruminski, J., & Grulkowski, I. (2026). Layer-specific quantitative analysis of accommodation in the human lens using long-range OCT and machine learning. Eye and Vision, 13(1), Article 47. https://doi.org/10.1186/s40662-026-00517-w

Image Credits: AI Generated

DOI: 10.1186/s40662-026-00517-w

Keywords: accommodation, crystalline lens, optical coherence tomography, machine learning, lens nucleus, lens cortex, OSD zones, biometry, presbyopia, eye anatomy, swept-source OCT, refraction correction