The Predictive Capabilities of Artificial Intelligence-Based OCT Analysis for Age-Related Macular Degeneration Progression—A Systematic Review
Abstract
:1. Introduction
2. Materials and Methods
2.1. Methodology of the Literature Search
ArXiv | 18 articles | (https://arxiv.org/) |
Cochrane Library | 18 articles | (https://www.cochranelibrary.com/) |
Embase | 258 articles | (https://www.embase.com) |
IEEE Xplore | 29 articles | (https://ieeexplore.ieee.org) |
Pubmed | 113 articles | (https://pubmed.ncbi.nlm.nih.gov/) |
Science Direct | 362 articles | (https://www.sciencedirect.com) |
Scopus | 1002 articles | (https://www.scopus.com/) |
2.2. Inclusion Criteria
- The progression of AMD from early and iAMD stages to nAMD and GA;
- The progression of nAMD under anti-VEGF therapy;
- The progression of GA in its natural history and under new trial treatments;
- The efficacy of anti-VEGF treatment of nAMD;
- The requirements of anti-VEGF treatment of nAMD;
- The VA outcomes after anti-VEGF treatment of nAMD.
2.3. Exclusion Criteria
- Articles that used ML and DL algorithms but based their prediction on color fundus photographs and other types of imaging (e.g., confocal scanning laser ophthalmoscopy) and did not include OCT analysis;
- Articles that analyzed OCT images with ML and DL algorithms but focused on learning tasks, such as classification or segmentation of AMD;
- Articles that analyzed OCT images with ML and DL models but used statistical analysis for their prediction models.
2.4. Selection of Papers
3. Results
- Four articles predicted the need for anti-VEGF treatment among nAMD patients; three used ML models and one used a DL model;
- Four articles predicted anti-VEGF treatment efficacy for nAMD patients; all four used DL models and, among them, two used generative adversarial networks (GANs);
- Four articles predicted the conversion from early or iAMD to nAMD; one used a ML model and three used DL models;
- Two articles predicted the conversion from iAMD to GA; one used a ML model and one used a DL model;
- Three articles predicted the growth of GA; one used a ML model and two used DL models;
- Three articles predicted the VA outcome after anti-VEGF treatment in nAMD patients; two used ML models and one used a DL model.
Prediction | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 |
---|---|---|---|---|---|---|
Treatment requirements | Bogunovic et al. [14] | Romo-Bucheli et al. [15] | Pfau et al. [16] | |||
Gallardo et al. [17] | ||||||
Treatment efficacy | Feng et al. [18]; | Lee et al. [19]; | ||||
Liu et al. [20] | Zhao et al. [21] | |||||
Conversion to nAMD | Schmidt-Erfurth et al. [13] | Russakoff et al. [22] | Yim et al. [23]; | |||
Banerjee et al. [24] | ||||||
Conversion to GA | Schmidt-Erfurth et al. [13] | Rivail et al. [25] | ||||
GA growth | Niu et al. [26] | Zhang et al. [27]; | ||||
Gigon et al. [28] | ||||||
VA outcome | Schmidt-Erfurth et al. [29] | Kawczynski et al. [30] | ||||
Rohm et al. [31] |
3.1. Prediction of Treatment Requirements—Classic Machine Learning
Lasso Regression | Principal Component Regression | Random Forest Regression | NGBoost | |
---|---|---|---|---|
MAE for no. of inj. for 12 months (95% CI) | 2.76 (2.39–3.14) ( = 0.038) | 2.74 (2.38–3.11) ( = 0.173) | 2.60 (2.25–2.96) ( = 0.390) | 2.66 (2.31–3.01) ( = 0.094) |
AUC for low Tx requirement (≤4 inj.) | 0.61 | 0.63 | 0.68 | 0.68 |
AUC for high Tx requirement (≥10 inj.) | 0.63 | 0.7 | 0.7 | 0.69 |
3.2. Prediction of Treatment Requirements—Deep Learning
3.3. Prediction of nAMD Anti-VEGF Treatment Efficacy—Deep Learning
3.4. Prediction of Conversion to nAMD and GA—Classic Machine Learning
3.5. Prediction of Conversion to nAMD—Deep Learning
3.6. Prediction of Conversion to GA—Deep Learning
3.7. Prediction of GA Growth—Classic Machine Learning
3.8. Prediction of GA Growth—Deep Learning
3.9. Prediction of VA Outcome—Classic Machine Learning
3.10. Prediction of VA Outcome—Deep Learning
4. Discussion
4.1. Prediction of Treatment Requirement
4.2. Prediction of Treatment Efficacy
Authors | Small Sample Size | Data Source | External Validation | Automatically Extracted Features | Quality of Efficacy Interpretation |
---|---|---|---|---|---|
Feng et al. [18] | No | RWD | No | N/A | Unknown |
Liu et. al. [20] | No | RWD | No | N/A | RLCP OCT biomarkers |
Lee et al. [19] | No | RWD | No | N/A | RLCP OCT biomarkers |
Zhao et al. [21] | No | RWD | No | N/A | RLCP BCVA |
4.3. Prediction of Conversion to nAMD
4.4. Prediction of Conversion to GA
4.5. Prediction of GA Growth
4.6. Prediction of VA Outcome
Authors | Date | Model | Dataset |
---|---|---|---|
Machine Learning | |||
Schmidt-Erfurth et al. [29] | January 2018 | Random forest regression | 614 nAMD eyes of 614 patients undergoing anti-VEGF treatment within the HARBOR clinical trial |
Rohm et al. [31] | July 2018 | AdaBoost.R2, gradient boosting, random forests, extremely randomized trees, lasso | 738 nAMD eyes from 653 patients (for 3-month forecast) and 508 nAMD eyes from 456 patients (for 12-month forecast) visiting Ludwig Maximilian University of Munich, Germany |
Deep Learning | |||
Kawczynski et al. [30] | September 2020 | ResNet-50 v2 | 1071 nAMD eyes of 1071 patients undergoing anti-VEGF treatment within the HARBOR clinical trial |
Author and Prediction | Model and Scenario | No. of Visits | Prediction Time | R2 | RMSE |
---|---|---|---|---|---|
Machine Learning | |||||
Schmidt-Erfurth et al. [29]—BCVA at 12 months | Random forest regression | BL + 3 V | 12 m | 0.70 | 8.6 letters |
Rohm et al. [31]—BCVA at 3 and 12 months | AdaBoost.R2 | BL + 3 V | 12 m | N/A | 0.2 logMAR |
AdaBoost.R2 | BL + 3 V | 3 m | N/A | 0.25 logMAR | |
Gradient boosting | BL + 3 V | 12 m | N/A | 0.19 logMAR | |
Gradient boosting | BL + 3 V | 3 m | N/A | 0.2 logMAR | |
Random forests | BL + 3 V | 12 m | N/A | 0.18 logMAR | |
Random forests | BL + 3 V | 3 m | N/A | 0.23 logMAR | |
Extremely randomized trees | BL + 3 V | 12 m | N/A | 0.18 logMAR | |
Extremely randomized trees | BL + 3 V | 3 m | N/A | 0.23 logMAR | |
Lasso | BL + 3 V | 12 m | N/A | 0.2 logMAR | |
Lasso | BL + 3 V | 3 m | N/A | 0.18 logMAR | |
Deep Learning | |||||
Kawczynski et al. [30].—BCVA at current visit and 12 months | ResNet-50 v2—study eyes | BL | 12 m | 0.33 | 14.16 letters |
ResNet-50 v2—fellow eyes | BL | 12 m | 0.75 | 11.27 letters |
Authors | Small Sample Size | Data Source | External Validation | Automatically Extracted Features | Quality of VA Measurement |
---|---|---|---|---|---|
Schmidt-Erfurth et al. [29] | No | RCT | No | Yes | RCT measurement |
Rohm et al. [31] | No | RWD | No | Yes | RCLP measurement |
Kawczynski et al. [30] | No | RCT | No | N/A | RCT measurement |
4.7. Quality Assurance for AI Models
4.8. Challenges for AI in Real-World Clinical Settings
5. Conclusions
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Conflicts of Interest
References
- Wong, W.L.; Su, X.; Li, X.; Cheung, C.M.G.; Klein, R.; Cheng, C.Y.; Wong, T.Y. Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: A systematic review and meta-analysis. Lancet Glob. Health 2014, 2, e106–e116. [Google Scholar] [PubMed]
- Khan, H.; Aziz, A.A.; Sulahria, H.; Khan, H.; Ahmed, A.; Choudhry, N.; Narayanan, R.; Danzig, C.; Khanani, A.M. Emerging Treatment Options for Geographic Atrophy (GA) Secondary to Age-Related Macular Degeneration. Clin. Ophthalmol. 2023, 17, 321–327. [Google Scholar] [PubMed]
- Heier, J.S.; Khanani, A.M.; Ruiz, C.Q.; Basu, K.; Ferrone, P.J.; Brittain, C.; Figueroa, M.S.; Lin, H.; Holz, F.G.; Patel, V.; et al. Efficacy, durability, and safety of intravitreal faricimab up to every 16 weeks for neovascular age-related macular degeneration (TENAYA and LUCERNE): Two randomised, double-masked, phase 3, non-inferiority trials. Lancet 2022, 399, 729–740. [Google Scholar] [PubMed]
- Maguire, M.G.; Daniel, E.; Shah, A.R.; Grunwald, J.E.; Hagstrom, S.A.; Avery, R.L.; Huang, J.; Martin, R.W.; Roth, D.B.; Castellarin, A.A.; et al. Incidence of choroidal neovascularization in the fellow eye in the comparison of age-related macular degeneration treatments trials. Ophthalmology 2013, 120, 2035–2041. [Google Scholar]
- Fujimoto, J.; Swanson, E. The development, commercialization, and impact of optical coherence tomography. Investig. Ophthalmol. Vis. Sci. 2016, 57, OCT1–OCT13. [Google Scholar]
- Schmidt-Erfurth, U.; Sadeghipour, A.; Gerendas, B.S.; Waldstein, S.M.; Bogunović, H. Artificial intelligence in retina. Prog. Retin. Eye Res. 2018, 67, 1–29. [Google Scholar]
- LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [Google Scholar]
- Abràmoff, M.D.; Lavin, P.T.; Birch, M.; Shah, N.; Folk, J.C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit. Med. 2018, 1, 39. [Google Scholar]
- Ho, A.C.; Albini, T.A.; Brown, D.M.; Boyer, D.S.; Regillo, C.D.; Heier, J.S. The potential importance of detection of neovascular age-related macular degeneration when visual acuity is relatively good. JAMA Ophthalmol. 2017, 135, 268–273. [Google Scholar]
- Chew, E.Y.; Clemons, T.E.; Bressler, S.B.; Elman, M.J.; Danis, R.P.; Domalpally, A.; Heier, J.S.; Kim, J.E.; Garfinkel, R.; Group, A.H.S.R.; et al. Randomized trial of a home monitoring system for early detection of choroidal neovascularization home monitoring of the Eye (HOME) study. Ophthalmology 2014, 121, 535–544. [Google Scholar]
- Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Akl, E.A.; Brennan, S.E.; et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. Int. J. Surg. 2021, 88, 105906. [Google Scholar] [CrossRef]
- Moher, D.; Tetzlaff, J.; Tricco, A.C.; Sampson, M.; Altman, D.G. Epidemiology and reporting characteristics of systematic reviews. PLoS Med. 2007, 4, e78. [Google Scholar] [CrossRef] [Green Version]
- Schmidt-Erfurth, U.; Waldstein, S.M.; Klimscha, S.; Sadeghipour, A.; Hu, X.; Gerendas, B.S.; Osborne, A.; Bogunović, H. Prediction of individual disease conversion in early AMD using artificial intelligence. Investig. Ophthalmol. Vis. Sci. 2018, 59, 3199–3208. [Google Scholar] [CrossRef] [PubMed] [Green Version]
- Bogunović, H.; Waldstein, S.M.; Schlegl, T.; Langs, G.; Sadeghipour, A.; Liu, X.; Gerendas, B.S.; Osborne, A.; Schmidt-Erfurth, U. Prediction of anti-VEGF treatment requirements in neovascular AMD using a machine learning approach. Investig. Ophthalmol. Vis. Sci. 2017, 58, 3240–3248. [Google Scholar] [CrossRef] [PubMed] [Green Version]
- Romo-Bucheli, D.; Erfurth, U.S.; Bogunović, H. End-to-end deep learning model for predicting treatment requirements in neovascular AMD from longitudinal retinal OCT imaging. IEEE J. Biomed. Health Inform. 2020, 24, 3456–3465. [Google Scholar] [CrossRef]
- Pfau, M.; Sahu, S.; Rupnow, R.A.; Romond, K.; Millet, D.; Holz, F.G.; Schmitz-Valckenberg, S.; Fleckenstein, M.; Lim, J.I.; de Sisternes, L.; et al. Probabilistic forecasting of anti-VEGF treatment frequency in neovascular age-related macular degeneration. Transl. Vis. Sci. Technol. 2021, 10, 30. [Google Scholar] [CrossRef] [PubMed]
- Gallardo, M.; Munk, M.R.; Kurmann, T.; De Zanet, S.; Mosinska, A.; Karagoz, I.K.; Zinkernagel, M.S.; Wolf, S.; Sznitman, R. Machine learning can predict anti-VEGF treatment demand in a treat-and-extend regimen for patients with neovascular AMD, DME, and RVO associated macular edema. Ophthalmol. Retin. 2021, 5, 604–624. [Google Scholar] [CrossRef]
- Feng, D.; Chen, X.; Zhou, Z.; Liu, H.; Wang, Y.; Bai, L.; Zhang, S.; Mou, X. A preliminary study of predicting effectiveness of anti-VEGF injection using OCT images based on deep learning. In Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020; IEEE: Piscataway, NJ, USA, 2020; pp. 5428–5431. [Google Scholar]
- Lee, H.; Kim, S.; Kim, M.A.; Chung, H.; Kim, H.C. Post-treatment prediction of optical coherence tomography using a conditional generative adversarial network in age-related macular degeneration. Retina 2021, 41, 572–580. [Google Scholar] [CrossRef]
- Liu, Y.; Yang, J.; Zhou, Y.; Wang, W.; Zhao, J.; Yu, W.; Zhang, D.; Ding, D.; Li, X.; Chen, Y. Prediction of OCT images of short-term response to anti-VEGF treatment for neovascular age-related macular degeneration using generative adversarial network. Br. J. Ophthalmol. 2020, 104, 1735–1740. [Google Scholar] [CrossRef]
- Zhao, X.; Zhang, X.; Lv, B.; Meng, L.; Zhang, C.; Liu, Y.; Lv, C.; Xie, G.; Chen, Y. Optical coherence tomography-based short-term effect prediction of anti-vascular endothelial growth factor treatment in neovascular age-related macular degeneration using sensitive structure guided network. Graefe’S Arch. Clin. Exp. Ophthalmol. 2021, 259, 3261–3269. [Google Scholar] [CrossRef]
- Russakoff, D.B.; Lamin, A.; Oakley, J.D.; Dubis, A.M.; Sivaprasad, S. Deep learning for prediction of AMD progression: A pilot study. Investig. Ophthalmol. Vis. Sci. 2019, 60, 712–722. [Google Scholar] [CrossRef] [PubMed] [Green Version]
- Yim, J.; Chopra, R.; Spitz, T.; Winkens, J.; Obika, A.; Kelly, C.; Askham, H.; Lukic, M.; Huemer, J.; Fasler, K.; et al. Predicting conversion to wet age-related macular degeneration using deep learning. Nat. Med. 2020, 26, 892–899. [Google Scholar] [CrossRef] [PubMed]
- Banerjee, I.; de Sisternes, L.; Hallak, J.A.; Leng, T.; Osborne, A.; Rosenfeld, P.J.; Gregori, G.; Durbin, M.; Rubin, D. Prediction of age-related macular degeneration disease using a sequential deep learning approach on longitudinal SD-OCT imaging biomarkers. Sci. Rep. 2020, 10, 15434. [Google Scholar] [CrossRef] [PubMed]
- Rivail, A.; Schmidt-Erfurth, U.; Vogl, W.D.; Waldstein, S.M.; Riedl, S.; Grechenig, C.; Wu, Z.; Bogunovic, H. Modeling disease progression in retinal OCTs with longitudinal self-supervised learning. In Proceedings of the International Workshop on Predictive Intelligence in MEdicine, Shenzhen, China, 13 October 2019; Springer: Berlin/Heidelberg, Germany, 2019; pp. 44–52. [Google Scholar]
- Niu, S.; de Sisternes, L.; Chen, Q.; Rubin, D.L.; Leng, T. Fully automated prediction of geographic atrophy growth using quantitative spectral-domain optical coherence tomography biomarkers. Ophthalmology 2016, 123, 1737–1750. [Google Scholar] [CrossRef] [PubMed]
- Zhang, Y.; Zhang, X.; Ji, Z.; Niu, S.; Leng, T.; Rubin, D.L.; Yuan, S.; Chen, Q. An integrated time adaptive geographic atrophy prediction model for SD-OCT images. Med. Image Anal. 2021, 68, 101893. [Google Scholar] [CrossRef] [PubMed]
- Gigon, A.; Mosinska, A.; Montesel, A.; Derradji, Y.; Apostolopoulos, S.; Ciller, C.; De Zanet, S.; Mantel, I. Personalized atrophy risk mapping in age-related macular degeneration. Transl. Vis. Sci. Technol. 2021, 10, 18. [Google Scholar] [CrossRef]
- Schmidt-Erfurth, U.; Bogunovic, H.; Sadeghipour, A.; Schlegl, T.; Langs, G.; Gerendas, B.S.; Osborne, A.; Waldstein, S.M. Machine learning to analyze the prognostic value of current imaging biomarkers in neovascular age-related macular degeneration. Ophthalmol. Retin. 2018, 2, 24–30. [Google Scholar] [CrossRef]
- Kawczynski, M.G.; Bengtsson, T.; Dai, J.; Hopkins, J.J.; Gao, S.S.; Willis, J.R. Development of deep learning models to predict best-corrected visual acuity from optical coherence tomography. Transl. Vis. Sci. Technol. 2020, 9, 51. [Google Scholar] [CrossRef]
- Rohm, M.; Tresp, V.; Müller, M.; Kern, C.; Manakov, I.; Weiss, M.; Sim, D.A.; Priglinger, S.; Keane, P.A.; Kortuem, K. Predicting visual acuity by using machine learning in patients treated for neovascular age-related macular degeneration. Ophthalmology 2018, 125, 1028–1036. [Google Scholar] [CrossRef]
- Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv 2014, arXiv:1409.1556. [Google Scholar]
- Sadda, S.R.; Guymer, R.; Holz, F.G.; Schmitz-Valckenberg, S.; Curcio, C.A.; Bird, A.C.; Blodi, B.A.; Bottoni, F.; Chakravarthy, U.; Chew, E.Y.; et al. Consensus definition for atrophy associated with age-related macular degeneration on OCT: Classification of atrophy report 3. Ophthalmology 2018, 125, 537–548. [Google Scholar] [CrossRef] [PubMed] [Green Version]
- Dong, L.; Yang, Q.; Zhang, R.H.; Wei, W.B. Artificial intelligence for the detection of age-related macular degeneration in color fundus photographs: A systematic review and meta-analysis. EClinicalMedicine 2021, 35, 100875. [Google Scholar] [CrossRef] [PubMed]
- Cheung, R.; Chun, J.; Sheidow, T.; Motolko, M.; Malvankar-Mehta, M.S. Diagnostic accuracy of current machine learning classifiers for age-related macular degeneration: A systematic review and meta-analysis. Eye 2022, 36, 994–1004. [Google Scholar] [CrossRef]
- Leng, X.; Shi, R.; Wu, Y.; Zhu, S.; Cai, X.; Lu, X.; Liu, R. Deep learning for detection of age-related macular degeneration: A systematic review and meta-analysis of diagnostic test accuracy studies. PLoS ONE 2023, 18, e0284060. [Google Scholar] [CrossRef]
- Seeböck, P.; Waldstein, S.M.; Klimscha, S.; Bogunovic, H.; Schlegl, T.; Gerendas, B.S.; Donner, R.; Schmidt-Erfurth, U.; Langs, G. Unsupervised identification of disease marker candidates in retinal OCT imaging data. IEEE Trans. Med. Imaging 2018, 38, 1037–1047. [Google Scholar] [CrossRef] [Green Version]
- Marginean, A.; Bianca, V.; Nicoara, S.D.; Muntean, G. Low-dimensional Representation of OCT Volumes with Supervised Contrastive Learning. In Proceedings of the 2022 IEEE 18th International Conference on Intelligent Computer Communication and Processing (ICCP), Cluj-Napoca, Romania, 22–24 September 2022; IEEE: Piscataway, NJ, USA, 2022; pp. 47–54. [Google Scholar]
- Patel, P.J.; Chen, F.K.; Rubin, G.S.; Tufail, A. Intersession repeatability of visual acuity scores in age-related macular degeneration. Investig. Ophthalmol. Vis. Sci. 2008, 49, 4347–4352. [Google Scholar] [CrossRef] [Green Version]
- Wiegand, T.; Lee, N.; Pujari, S.; Singh, M.; Xu, S.; Kuglitsch, M.; Lecoultre, M.; Riviere-Cinnamond, A.; Weicken, E.; Wenzel, M.; et al. Whitepaper for the ITU/WHO Focus Group on Artificial Intelligence for Health. 2019. Available online: https://www.itu.int/en/ITU-T/focusgroups/ai4h/Documents/FG-AI4H_Whitepaper.pdf (accessed on 10 January 2023).
- Oala, L.; Murchison, A.G.; Balachandran, P.; Choudhary, S.; Fehr, J.; Leite, A.W.; Goldschmidt, P.G.; Johner, C.; Schörverth, E.D.; Nakasi, R.; et al. Machine learning for health: Algorithm auditing & quality control. J. Med. Syst. 2021, 45, 105. [Google Scholar]
- Oala, L.; Fehr, J.; Gilli, L.; Balachandran, P.; Leite, A.W.; Calderon-Ramirez, S.; Li, D.X.; Nobis, G.; Alvarado, E.A.M.; Jaramillo-Gutierrez, G.; et al. Ml4h auditing: From paper to practice. Proc. Mach. Learn. Health PMLR 2020, 136, 280–317. [Google Scholar]
- Yanagihara, R.T.; Lee, C.S.; Ting, D.S.W.; Lee, A.Y. Methodological challenges of deep learning in optical coherence tomography for retinal diseases: A review. Transl. Vis. Sci. Technol. 2020, 9, 11. [Google Scholar] [CrossRef] [Green Version]
- Khan, S.M.; Liu, X.; Nath, S.; Korot, E.; Faes, L.; Wagner, S.K.; Keane, P.A.; Sebire, N.J.; Burton, M.J.; Denniston, A.K. A global review of publicly available datasets for ophthalmological imaging: Barriers to access, usability, and generalisability. Lancet Digit. Health 2021, 3, e51–e66. [Google Scholar] [CrossRef]
- Cabitza, F.; Campagner, A.; Albano, D.; Aliprandi, A.; Bruno, A.; Chianca, V.; Corazza, A.; Di Pietto, F.; Gambino, A.; Gitto, S.; et al. The elephant in the machine: Proposing a new metric of data reliability and its application to a medical case to assess classification reliability. Appl. Sci. 2020, 10, 4014. [Google Scholar] [CrossRef]
- Szegedy, C.; Zaremba, W.; Sutskever, I.; Bruna, J.; Erhan, D.; Goodfellow, I.; Fergus, R. Intriguing properties of neural networks. arXiv 2013, arXiv:1312.6199. [Google Scholar]
- Lynch, S.K.; Shah, A.; Folk, J.C.; Wu, X.; Abramoff, M.D. Catastrophic failure in image-based convolutional neural network algorithms for detecting diabetic retinopathy. Investig. Ophthalmol. Vis. Sci. 2017, 58, 3776. [Google Scholar]
- Nguyen, A.; Yosinski, J.; Clune, J. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 427–436. [Google Scholar]
- Gunning, D.; Stefik, M.; Choi, J.; Miller, T.; Stumpf, S.; Yang, G.Z. XAI—Explainable artificial intelligence. Sci. Robot. 2019, 4, 37. [Google Scholar] [CrossRef] [PubMed] [Green Version]
AUC | Accuracy | Sensitivity | Specificity | |
---|---|---|---|---|
Full w DA | 0.81 | 0.72 | 0.78 | 0.71 |
Full w/o DA | 0.84 | 0.77 | 0.98 | 0.18 |
Lesion w DA | 0.83 | 0.75 | 0.82 | 0.63 |
Lesion w/o DA | 0.73 | 0.74 | 0.96 | 0.13 |
GA Area | Scenario One | Scenario Two | Scenario Three |
---|---|---|---|
Mean DI +/− SD | |||
Total GA | 0.81 +/− 0.12 | 0.84 +/− 0.10 | 0.87 +/− 0.06 |
Growth-only GA | 0.72 +/− 0.18 | 0.74 +/− 0.17 | 0.72 +/− 0.22 |
Scenario | Predicted GA Growth | Training Data | DI | CC |
---|---|---|---|---|
1 | All consecutive follow-up visits for each patient | First two follow-up visits from all other patients | 0.86 | 0.83 |
2 | All consecutive follow-up visits for each patient | All consecutive follow-up visits from all other patients | 0.89 | 0.84 |
3 | Third and subsequent follow-up visits for each patient | First two follow-up visits from the same patient | 0.89 | 0.86 |
4 | Last follow-up visit for each patient | All prior, consecutive follow-up visits from the same patient | 0.92 | 0.88 |
5 | Third and subsequent follow-up visits for all patient | First two follow-up visits from all patients | 0.88 | 0.85 |
6 | Last follow-up visit for all patients | All prior, consecutive follow-up visits from all patients | 0.90 | 0.86 |
RORA Area | Manually Segmented GT | Automatically Segmented GT | ||
---|---|---|---|---|
Mean DI | Mean DI | |||
Scenario A | Scenario B | Scenario A | Scenario B | |
Total RORA | 0.73 to 0.80 | 0.83 to 0.88 | 0.74 to 0.81 | 0.84 to 0.89 |
Growth-only RORA | 0.46 to 0.72 | 0.44 to 0.64 | 0.39 to 0.71 | 0.35 to 0.62 |
Predicted | <69 Letters | <59 Letters | <38 Letters | |
---|---|---|---|---|
BCVA at CCV | ||||
Study eyes | = 0.67 (RMSE = 8.60) | AUC = 0.89 | AUC = 0.92 | AUC = 0.92 |
Fellow eyes | = 0.84 (RMSE = 9.01) | AUC = 0.93 | AUC = 0.97 | AUC = 0.98 |
BCVA at 12 m | ||||
Study eyes | = 0.33 (RMSE = 14.16) | AUC = 0.80 | AUC = 0.84 | AUC = 0.77 |
Fellow eyes | = 0.75 (RMSE = 11.27) | AUC = 0.92 | AUC = 0.93 | AUC = 0.96 |
Article | Records Identified | Studies Included |
---|---|---|
The current systematic review | 1404 * | 19 |
Artificial intelligence for the detection of age-related macular degeneration in color fundus photographs: A systematic review and meta-analysis [34] | 632 * | 19 |
Diagnostic accuracy of current machine learning classifiers for age-related macular degeneration: A systematic review and meta-analysis [35] | 423 * | 14 |
Deep learning for detection of age-related macular degeneration: A systematic review and meta-analysis of diagnostic test accuracy studies [36] | 359 ** | 18 |
Authors | Date | Model | Dataset |
---|---|---|---|
Machine Learning | |||
Bogunovic et al. [14] | June 2017 | Random forest classifier | 317 eyes from 317 nAMD patients undergoing PRN treatment within the HARBOR clinical trial |
Pfau et al. [16] | June 2021 | NGBoost | 40 eyes from 37 nAMD patients visiting the Department of Ophthalmology and Visual Sciences, University of Illinois, Chicago, and 59 nAMD eyes from 59 patients visiting the University Eye Hospital Bonn, University of Bonn, Germany |
Gallardo et al. [17] | July 2021 | Random forest classifier | 377 eyes from 340 nAMD patients undergoing T&E treatment at the University Hospital of Bern |
Deep Learning | |||
Romo Bucheli et al. [15] | June 2020 | DenseNet + RNN trained end to end | 350 nAMD patients undergoing PRN treatment |
Author and Prediction Task | Time Frame | Anti-VEGF Protocol | Model and Scenario | Number of Visits | AUC | Acc. |
---|---|---|---|---|---|---|
Machine Learning | ||||||
Bogunovic et al. [14]—low and high anti-VEGF inj. requirements in RCT (no. of inj.) | 2 years | PRN | Low-Tx (≤5 inj.) group vs. others | BL + 2 V | 0.7 | N/A |
High-Tx group (≥16 inj.) vs. others | BL + 2 V | 0.77 | N/A | |||
Pfau et al. [16]—low and high anti-VEGF inj. requirements in RWD (no. of inj.) | 1 year | PRN and T&E | Low-Tx group (≤4 inj.) | BL | 0.68 | N/A |
High-Tx group (≥10 inj.) | BL | 0.69 | N/A | |||
Gallardo et al. [17]—low and high anti-VEGF inj. requirements in RWD (interval between inj.) | 1 year | T&E | Low-Tx group (≥10 weeks) vs. others | BL + 2 V | 0.79 | N/A |
High-Tx group (≤5) vs. others | BL + 2 V | 0.79 | N/A | |||
Deep Learning | ||||||
Romo Bucheli et al. [15]—Low, medium and high anti-VEGF inj. requirements (no. of inj.) | 2 year | PRN | Low-Tx group (≤5) | BL + 2 V | N/A | 0.72 |
Intermediate-Tx group (>5 and <16) | BL + 2 V | N/A | 0.65 | |||
High-Tx group (≥ 16) | BL + 2 V | N/A | 0.9 | |||
Low-Tx group (≤5) vs. others | BL + 2 V | 0.85 | N/A | |||
High-Tx group (≥16) vs. others | BL + 2 V | 0.81 | N/A |
Authors | Small Sample Size | Data Source | External Validation | Automatically Extracted Features | Quality of Treatment Decision |
---|---|---|---|---|---|
Bogunovic et al. [14] | No | RCT | No | Yes | RCT protocol |
Pfau et al. [16] | Yes | RWD | No | Yes | RCLP protocol |
Gallardo et al. [17] | No | RWD | No | Yes | RCLP protocol |
Romo Bucheli et al. [15] | No | Unknown | No | Yes | RCLP protocol * |
Authors | Date | Model | Dataset |
---|---|---|---|
Feng et al. [18] | August 2020 | ResNet-50 | 228 patients with CNV, CME or both treated with anti-VEGF at Second Affiliated Hospital of Xi’an Jiaotong University between 2017 and 2019 |
Liu et al. [20] | March 2020 | pix2pixHD GAN | 526 OCT images from nAMD patients treated with anti-VEGF at Peking Union Medical College Hospital between 2018 and 2019 |
Lee et al. [19] | March 2021 | cGAN | 309 eyes of 298 nAMD patients treated with anti-VEGF at Konkuk University Medical Center between 2010 and 2019 |
Zhao et al. [21] | June 2021 | SSG-Net | 206 eyes of 181 nAMD patients treated with anti-VEGF at Peking Union Medical College Hospital between 2018 and 2020 |
Study | Input | Learning Task |
---|---|---|
Machine learning—Prediction of treatment requirements | ||
Bogunovic et al. [14] | Quantitative spatiotemporal features extracted from OCT images (at baseline and months one and two) + BCVA + demographic features + FA pattern type + smoking status | Low and high anti-VEGF injection requirements during 2-year PRN schedule in RCT |
Gallardo et al. [17] | Quantitative spatiotemporal features extracted from OCT (at baseline and months one and two) + demographic features | Low and high anti-VEGF injection requirements during 1-year T&E schedule in RWCP |
Pfau et al. [16] | Quantitative spatiotemporal features extracted from OCT at baseline | Treatment frequency during 1 year and low and high anti-VEGF injection requirements during 1-year PRN and T&E schedule in RWCP |
Deep learning—Prediction of treatment requirements | ||
Romo Bucheli et al. [15] | OCT volumes (at baseline and months one and two) | Low, medium and high anti-VEGF injection requirements during 2-year PRN schedule |
Deep learning—Prediction of treatment efficacy | ||
Feng et al. [18] | Two-dimensional (2D) pre-therapeutic OCT image (full OCT image/lesion region) | Efficacy of anti-VEGF injection for CNV, CME or both at 21 days (3 weeks) |
Liu et al. [20] | Two-dimensional (2D) pre- and post-therapeutic OCT image | Post-therapeutic generated image’s capacity to predict the macular status as either wet or dry and macular wet-to-dry conversion after anti-VEGF |
Lee et al. [19] | Two-dimensional (2D) pre- and post-therapeutic OCT image and FA and ICGA image | Post-therapeutic generated image’s capacity to predict presence of IRF, SRF, PED and SHRM 30 days after anti-VEGF loading phase |
Zhao et al. [21] | Two-dimensional (2D) pre- and post-therapeutic OCT image | Efficacy of anti-VEGF injection for post-treatment BCVA at approx. 4 weeks |
Machine learning—Prediction of conversion to late exudative AMD | ||
Schmidt-Erfurth et al. [13] | Quantitative spatiotemporal features extracted from OCT (at baseline and months one, two, three and four) + demographic features + smoking status + genetic features (SNPs) | Risk for conversion from iAMD to nAMD and GA in the following 2 years |
Deep learning—Prediction of conversion to late exudative AMD | ||
Russakoff et al. [22] | OCT B-scans | Risk for conversion from early or iAMD to nAMD in the following 2 years |
Banerjee et al. [24] | Quantitative spatiotemporal features extracted from OCT + age + gender + race + smoking status + VA | Risk for conversion from early or iAMD to nAMD in the following 3, 6, 9, 12, 15, 18 and 21 months |
Quantitative spatiotemporal features extracted from OCT + age + gender + VA | Risk for conversion from early or iAMD to nAMD in the following 3, 6, 9, 12, 15, 18 and 21 months | |
Yim et al. [23] | Three-dimensional (3D) raw OCT image + 3D one-hot-encoded OCT segmentation map | Risk for conversion from non-nAMD to nAMD in the following 6 months |
Machine learning—Prediction of conversion to GA | ||
Schmidt-Erfurth et al. [13] | Quantitative spatiotemporal features extracted from OCT (at baseline and months one, two, three and four) + demographic features + smoking status + genetic (SNPs) | Risk for conversion from iAMD to nAMD and GA in the following 2 years |
Deep learning—Prediction of conversion to GA | ||
Rivail et al. [25] | OCT B-scans | Risk for conversion from iAMD to GA in the following 6, 12 and 18 months |
Machine learning—Prediction of GA growth | ||
Niu et al. [26] | Quantitative spatiotemporal features extracted from OCT at baseline and first follow-up visit in three different scenarios | Future GA growth regions in three different scenarios (see text for details) |
Deep learning—Prediction of GA growth | ||
Zhang et al. [27] | OCT volumes + time factors at baseline and first two sequential follow-up visits | Future GA growth regions in six different scenarios |
Gigon et al. [28] | Thirteen en-face maps: six layer-thickness maps with their six corresponding reflectance maps + one drusen height map (scenario A: from baseline OCT; scenario B: from preceding OCT) | Future RORA growth regions (scenario A: from baseline to future visits; scenario B: from preceding to next visit) |
Machine learning—Prediction of VA outcome | ||
Schmidt-Erfurth et al. [29] | Quantitative spatiotemporal features extracted from OCT (at baseline and months one, two and three) + BCVA + anti-VEGF dose and regimen | BCVA at 12 months |
Rohm et al. [31] | Quantitative spatiotemporal features extracted from OCT XML file (at baseline and months one, two and three) + BCVA + 40 clinical features from EMR | BCVA at 3 and 12 months |
Deep learning—Prediction of VA outcome | ||
Kawczynski et al. [30] | SD-OCT volumes | BCVA at current visit and at 12 months; BCVA < threshold (69, 59, 38 letters) at current visit and at 12 months |
Authors and Prediction Task | Scenarios | AUC | Acc. | Se. | Sp. | PPV | NPV |
---|---|---|---|---|---|---|---|
Feng et al. [18]—efficacy of anti-VEGF injection for CNV, CME or both at 21 days (3 weeks) | Full w DA | 0.81 | 0.72 | 0.78 | 0.71 | N/A | N/A |
Full w/o DA | 0.84 | 0.77 | 0.98 | 0.18 | N/A | N/A | |
Lesion w DA | 0.83 | 0.75 | 0.82 | 0.63 | N/A | N/A | |
Lesion w/o DA | 0.73 | 0.74 | 0.96 | 0.13 | N/A | N/A | |
Liu et al. [20]—post-therapeutically generated image’s capacity to predict the macular status as either wet or dry and macular wet-to-dry conversion after anti-VEGF | Macular status (wet/dry) | N/A | 0.85 | 0.84 | 0.86 | 0.88 | 0.82 |
Wet-to-dry conversion | N/A | 0.81 | 0.83 | 0.79 | 0.75 | 0.86 | |
Lee et al. [19]—post-therapeutically generated image’s capacity to predict the presence of IRF, SRF, PED and SHRM 30 days after anti-VEGF loading phase | IRF (OCT only) IRF (OCT + FA + ICGA) | N/A | 89.6 | 33.3 | 95.1 | 40.0 | 93.6 |
N/A | 92.6 | 33.3 | 98.4 | 66.7 | 93.8 | ||
SRF (OCT only) SRF (OCT + FA + ICGA) | N/A | 77.0 | 21.2 | 95.1 | 58.3 | 78.9 | |
N/A | 80.7 | 24.2 | 99.0 | 88.9 | 80.2 | ||
PED (OCT only) PED (OCT + FA + ICGA) | N/A | 77.0 | 70.4 | 94.6 | 97.2 | 54.7 | |
N/A | 80.7 | 74.5 | 97.3 | 98.7 | 59.0 | ||
SHRM (OCT only) SHRM (OCT + FA + ICGA) | N/A | 91.9 | 76.5 | 94.1 | 65.0 | 96.5 | |
N/A | 96.3 | 88.2 | 97.5 | 83.3 | 98.3 | ||
Zhao et al. [21]—efficacy of anti-VEGF injection on post treatment BCVA at approx. 28 days by classyifing responders and non-responders | Responders/non- responders | 0.83 | 84.6 | 0.692 | 1 | N/A | N/A |
Authors | Date | Model | Dataset |
---|---|---|---|
Machine Learning | |||
Schmidt-Erfurth et al. [13] | July 2018 | Cox proportional hazard | 495 iAMD fellow eyes of 495 nAMD patients undergoing anti-VEGF treatment within the HARBOR clinical trial |
Deep Learning | |||
Russakoff et al. [22] | February 2019 | AMDnet and VGG16 | 71 early or iAMD fellow eyes of 71 nAMD patients undergoing anti-VEGF treatment at Moorfields Eye Hospital |
Banerjee et al. [24] | September 2020 | Deep Sequence LSTM | 671 early or iAMD fellow eyes from 671 nAMD patients undergoing anti-VEGF treatment with the HARBOR clinical trial |
Deep Sequence LSTM | 719 early or iAMD eyes from 507 patients visiting Bascom Palmer Eye Institute (MIAMI) between 2004 to 2015 | ||
Yim et al. [23] | May 2020 | Ensemble DL model | 2261 non-nAMD fellow eyes from 2795 nAMD patients visiting seven sites of Moorfields Eye Hospital between 2012 and 2017 |
Author and Prediction Task | Model and Scenario | Number of Visits | Prediction Time Frame | AUC | Se. | Sp. |
---|---|---|---|---|---|---|
Schmidt-Erfurth et al. [13]—risk for conversion from iAMD to nAMD | CPH | BL + 4 V | 24 m | 0.68 | 0.80 | 0.46 |
Russakoff et al. [22]—risk for conversion from early or iAMD to nAMD | AMDNet—OCT B-scan | BL | 24 m | 0.89 | N/A | N/A |
AMDnet—OCT volume | BL | 24 m | 0.91 | N/A | N/A | |
VGG16—OCT B-scan | BL | 24 m | 0.82 | N/A | N/A | |
VGG16—OCT volume | BL | 24 m | 0.87 | N/A | N/A | |
Banerjee et al. [24]—risk for conversion from early or iAMD to nAMD | Deep Sequence LSTM (Harbor) | TA | 3 m | 0.96 | N/A | N/A |
TA | 6 m | 0.83 | N/A | N/A | ||
TA | 9 m | 0.78 | N/A | N/A | ||
TA | 12 m | 0.77 | N/A | N/A | ||
TA | 15 m | 0.84 | N/A | N/A | ||
TA | 18 m | 0.90 | N/A | N/A | ||
TA | 21 m | 0.97 | N/A | N/A | ||
BL + 4 V | 18 m | N/A | 0.88 | 0.87 | ||
Deep Sequence LSTM (Miami) | TA | 3 m | 0.82 | N/A | N/A | |
TA | 6 m | 0.77 | N/A | N/A | ||
TA | 9 m | 0.71 | N/A | N/A | ||
TA | 12 m | 0.69 | N/A | N/A | ||
TA | 15 m | 0.68 | N/A | N/A | ||
TA | 18 m | 0.65 | N/A | N/A | ||
TA | 21 m | 0.68 | N/A | N/A | ||
BL + 4 V | 18 m | N/A | 0.68 | 0.71 | ||
Yim et al. [23]—risk for conversion from non-nAMD to nAMD | Ensemble DL model—OCT volume (conversion scan) | One | 6 m | 0.745 | 0.80 (L); 0.34 (C) | 0.55 (L); 0.90 (C) |
Ensemble DL model—OCT volume (injection scan) | One | 6 m | 0.884 | 0.80 (L); 0.34 (C) | 0.55 (L); 0.90 (C) |
Authors | Small Sample Size | Data Source | External Validation | Automatically Extracted Features | Quality of Conversion Interpretation |
---|---|---|---|---|---|
Schmidt-Erfurth et al. [13] | No | RCT | No | Yes | RCT OCT monitoring |
Russakoff et al. [22] | Yes | RWD | No | N/A | RLCP OCT and FA monitoring |
Banerjee et al. [24] | No | RCT | Yes | Yes | RCT OCT monitoring |
Yim et al. [23] | No | RWD | No | Yes | RLCP OCT monitoring and injection start |
Authors | Date | Dataset | Model |
---|---|---|---|
Machine Learning | |||
Schmidt-Erfurth et al. [13] | July 2018 | 495 iAMD fellow eyes of 495 nAMD patients undergoing anti-VEGF treatment within the HARBOR clinical trial | Cox proportional hazard |
Deep Learning | |||
Rivail et al. [25] | October 2019 | 420 iAMD eyes from 221 patients | Deep Siamese network |
Author and Prediction | Model | No. of Visits | Time Frame | AUC | Se | Sp | P |
---|---|---|---|---|---|---|---|
Machine Learning | |||||||
Schmidt-Erfurth et al. [13]—risk for conversion from iAMD to GA | CPH | BL + 4 V | 24 m | 0.80 | 0.80 | 0.69 | N/A |
Deep Learning | |||||||
Rivail et al. [25]—risk for conversion from iAMD to GA | Deep Siamese network | One | 6 m | 0.753 | N/A | N/A | 0.367 |
One | 12 m | 0.784 | N/A | N/A | 0.394 | ||
One | 18 m | 0.773 | N/A | N/A | 0.463 |
Authors | Small Sample Size | Data Source | External Validation | Automatically Extracted Features | Quality of Conversion Interpretation |
---|---|---|---|---|---|
Schmidt-Erfurth et al. [13] | No | RCT | No | Yes | RCT OCT monitoring |
Rivail et al. [25] | No | Unknown | No | N/A | Unknown |
Authors | Date | Model | Dataset |
---|---|---|---|
Machine Learning | |||
Niu et al. [26] | June 2016 | Random forest classifier | 38 GA eyes from 29 patients visiting Byers Eye Institute of Stanford University |
Deep Learning | |||
Zhang et al. [27] | February 2021 | BiLSTM + 3D-UNet | 22 GA eyes from 22 patients visiting Byers Eye Institute of Stanford University and 3 GA eyes from 3 patients visiting Jiangsu Provincial People’s Hospital in China |
Gigon et al. [28] | November 2021 | Encoder–decoder CNN + time-based Taylor series | 129 GA eyes from 119 patients who visited the Jules Gonin Eye Hospital in Lausanne, Switzerland |
Scenario | Predicted GA Growth | Training Data | DI |
---|---|---|---|
Niu et al. [26]—Random Forest Classifier | |||
1 | First follow-up visit for each patient | BL + first follow-up from all other patients | 0.81 * |
2 | All consecutive follow-up visits for each patient | BL + first follow-up from all other patients | 0.84 * |
3 | Second and subsequent follow-up visits for each patient | BL + first follow-up scan from the same patient | 0.87 * |
Zhang et al. [27]—BiLSTM + 3D-UNet | |||
1 | All consecutive follow-up visits for each patient | BL + first two follow-up visits from all other patients | 0.86 * |
2 | All consecutive follow-up visits for each patient | All consecutive follow-up visits from all other patients | 0.89 * |
3 | Third and subsequent follow-up visits for each patient | BL + first two follow-up visits from the same patient | 0.89 * |
4 | Last follow-up visit for each patient | All prior consecutive follow-up visits from the same patient | 0.92 * |
5 | Third and subsequent follow-up visits for all patients | BL + first two follow-up visits from all patients | 0.88 * |
6 | Last follow-up visit for all patients | All prior consecutive follow-up visits from all patients | 0.90 * |
Gigon et al. [28]—Encoder–Decoder CNN + Time-based Taylor Series | |||
A | BL and subsequent follow-up visits | BL visit—manually segmented GT | 0.73 to 0.80 * |
A | BL and subsequent follow-up visits | BL visit—automatically segmented GT | 0.74 to 0.81 * |
B | Next follow-up visit | Preceding visit—manually segmented GT | 0.83 to 0.88 * |
B | Next follow-up visit | Preceding visit—automatically segmented GT | 0.84 to 0.89 * |
Authors | Small Sample Size | Data Source | External Validation | Automatically Extracted Features | Quality of GA Growth GT |
---|---|---|---|---|---|
Niu et al. [26] | Yes | RWD | No | Yes | Automatic segmentation with visual review |
Zhang et al. [27] | Yes | RWD | No | N/A | Automatic segmentation with visual review |
Gigon et al. [28] | No | RWD | No | Yes | Automatic and manual segmentation |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Muntean, G.A.; Marginean, A.; Groza, A.; Damian, I.; Roman, S.A.; Hapca, M.C.; Muntean, M.V.; Nicoară, S.D. The Predictive Capabilities of Artificial Intelligence-Based OCT Analysis for Age-Related Macular Degeneration Progression—A Systematic Review. Diagnostics 2023, 13, 2464. https://doi.org/10.3390/diagnostics13142464
Muntean GA, Marginean A, Groza A, Damian I, Roman SA, Hapca MC, Muntean MV, Nicoară SD. The Predictive Capabilities of Artificial Intelligence-Based OCT Analysis for Age-Related Macular Degeneration Progression—A Systematic Review. Diagnostics. 2023; 13(14):2464. https://doi.org/10.3390/diagnostics13142464
Chicago/Turabian StyleMuntean, George Adrian, Anca Marginean, Adrian Groza, Ioana Damian, Sara Alexia Roman, Mădălina Claudia Hapca, Maximilian Vlad Muntean, and Simona Delia Nicoară. 2023. "The Predictive Capabilities of Artificial Intelligence-Based OCT Analysis for Age-Related Macular Degeneration Progression—A Systematic Review" Diagnostics 13, no. 14: 2464. https://doi.org/10.3390/diagnostics13142464