Related Experiment Video
Updated: Aug 27, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Target-label-free artificial intelligence framework for cross-anatomical RNFL biomarker segmentation in optical
Suo Qiu1,2, Juntao Zhang1,2, Na Zhao1,2
1The Affiliated People's Hospital of Ningbo University, The Eye Hospital of Wenzhou Medical University (Ningbo Branch), Ningbo, China.
Background:
Effective glaucoma management relies on accurate optical coherence tomography (OCT) quantification of the circumpapillary retinal nerve fiber layer (cpRNFL) to track disease progression. However, the clinical translation of artificial intelligence (AI)-driven structural assessment is bottlenecked by scarce optic nerve head (ONH) annotations and coupled anatomical-hardware domain shifts.
Methods:
We propose Macula2Disc (M2D), a target-label-free, transductive AI framework for cross-anatomical image analysis. Trained exclusively on source-domain macular B-scans, M2D bridges the macula-to-disc gap across heterogeneous OCT platforms (Heidelberg Spectralis and TowardPi BMizar-400K) via three integrated components: (1) a Three-Level Deformation Module (TLDM) synthesizing ONH morphological variations from warp-stable macular topologies; (2) a Non-Uniform Rational B-Spline (NURBS) photometric module calibrating device-specific intensity manifolds; and (3) a Cross-Domain Segmentation Network (CDSN) optimized through entropy-gated global feature alignment.
Results:
Validation on 1,017 unannotated ONH circumpapillary scans (422 participants) yielded an overall RNFL Dice coefficient of 87.42% and an MIoU of 80.43%. Evaluated against expert consensus in a mixed-device clinical subset, M2D demonstrated statistical concordance (Mean Absolute Difference [MAD] = 1.8 m, ICC(2,1) = 0.96, ). In cases of pronounced RNFL attenuation, an advanced pathological state where proprietary commercial software exhibited statistically significant boundary-tracking deviations (MAD = 2.8 m, ), the proposed framework achieved anatomical alignment consistent with expert consensus. Across the entire cohort, M2D structural biomarker measurements demonstrated statistical agreement with commercial baseline outputs (Overall MAE = 1.5 m), indicating a robust technical capacity to approximate proprietary segmentations across different hardware architectures.
Conclusion:
The M2D framework provides a target-label-free AI strategy to resolve coupled macula-to-disc shifts. By approximating commercial anatomical measurements in general populations while resolving boundary-tracking vulnerabilities in pathological extremes, this approach establishes a foundational computational basis for cross-platform anatomical assessment. The primary endpoint of this study is the consistent technical evaluation of structural biomarker extraction, with rigorous diagnostic performance analysis remaining a crucial future direction.
