Multimodal Cataract Severity Classification from AS-OCT Examination Reports: Analyzing the Complementarity of B-Scan Images and A-Scan Profiles
- 1Department of Data Science, , Seoul 01811, Republic of Korea; (M.-S.K.); (G.-M.P.)
- 2Department of Ophthalmology, , Seoul 04564, Republic of Korea
Abstract
Background/Objectives: Deep learning on anterior segment optical coherence tomography (AS-OCT) has been applied mainly to nuclear cataract. This study evaluated whether the A-scan profile in Anterion reports adds information to the B-scan image for classifying nuclear (N), cortical (C), and posterior subcapsular (P) cataract severity. Methods: B-scan images and A-scan profiles were extracted from 1416 reports of 691 patients. LOCS III grades were converted into three levels (None/Mild, Moderate, Advanced); inter-observer agreement was excellent (weighted κ 0.84–0.97). Image and profile encoders were combined using five fusion strategies and evaluated with patient-level 5-fold cross-validation. Results: Adding the A-scan profile improved F1 for N (+0.030) and P (+0.094) but not for C (−0.002). All fusion strategies outperformed single-modality models and performed comparably (average F1 0.514–0.531). For P, the gain was largest at the Moderate grade (F1 0.183 to 0.336), and masking the profile segment containing the posterior capsule caused the largest drop. B-scan line profiles could not replace the device-provided profile, and most misclassifications occurred between adjacent grades. Conclusions: The A-scan profile provided a useful incremental contribution, particularly for N and P, in this single-center dataset. The approach may be useful in retrospective and multi-center settings, but this remains to be demonstrated.
Related
- Toward integrated multimodal artificial intelligence in bone marrow examination: current state and open challenges
- Expert-Guided Visual Correction for Characterizing Diagnostic Performance and Error Patterns of Multimodal Large Language Models Using Periodontal In-Service Examination Images
- Bilateral Long-Lasting Changes in Multimodal Sensitivity After Cutaneous High-Frequency Electrical Stimulation: A New Comprehensive Dermatomal Examination
- Multimodal patch-based deep learning for early Alzheimer’s disease classification using MRI and PET
- Multimodal fusion of SpO₂ time series and clinical data for explainable classification of obstructive sleep apnea severity
BibTeX
@article{kim2026multimodal,
title = {{Multimodal Cataract Severity Classification from AS-OCT Examination Reports: Analyzing the Complementarity of B-Scan Images and A-Scan Profiles}},
author = {Min-Seok Kim and Gang-Min Park and Ho-Gil Jung and Jong-Joo Lee},
journal = {Diagnostics (Basel, Switzerland)},
year = {2026},
month = oct,
doi = {10.3390/diagnostics16203362},
url = {https://europepmc.org/article/PMC/PMC13703474},
note = {CC BY},
}
RIS
TY - JOUR
TI - Multimodal Cataract Severity Classification from AS-OCT Examination Reports: Analyzing the Complementarity of B-Scan Images and A-Scan Profiles
AU - Min-Seok Kim
AU - Gang-Min Park
AU - Ho-Gil Jung
AU - Jong-Joo Lee
PY - 2026
DA - 2026/10/01
JO - Diagnostics (Basel, Switzerland)
DO - 10.3390/diagnostics16203362
UR - https://europepmc.org/article/PMC/PMC13703474
AB - Background/Objectives: Deep learning on anterior segment optical coherence tomography (AS-OCT) has been applied mainly to nuclear cataract. This study evaluated whether the A-scan profile in Anterion reports adds information to the B-scan image for classifying nuclear (N), cortical (C), and posterior subcapsular (P) cataract severity. Methods: B-scan images and A-scan profiles were extracted from 1416 reports of 691 patients. LOCS III grades were converted into three levels (None/Mild, Moderate, Advanced); inter-observer agreement was excellent (weighted κ 0.84–0.97). Image and profile encoders were combined using five fusion strategies and evaluated with patient-level 5-fold cross-validation. Results: Adding the A-scan profile improved F1 for N (+0.030) and P (+0.094) but not for C (−0.002). All fusion strategies outperformed single-modality models and performed comparably (average F1 0.514–0.531). For P, the gain was largest at the Moderate grade (F1 0.183 to 0.336), and masking the profile segment containing the posterior capsule caused the largest drop. B-scan line profiles could not replace the device-provided profile, and most misclassifications occurred between adjacent grades. Conclusions: The A-scan profile provided a useful incremental contribution, particularly for N and P, in this single-center dataset. The approach may be useful in retrospective and multi-center settings, but this remains to be demonstrated.
SN - 2075-4418
ER -