Deep-Learning-Assisted Analysis of Early Biomarker Changes in Treatment-Naïve Patients with Neovascular AMD Under Intravitreal Faricimab

This retrospective study screened the Smart Eye Database from the Department of Ophthalmology at LMU University Hospital Munich for patients treated with faricimab for nAMD between January 2023 and December 2024. The inclusion criteria were: (i) diagnosis of nAMD; (ii) no previous intravitreal or other treatment for nAMD at the study eye (treatment naïve); (iii) four loading doses of faricimab in a 3 ± 1 month time period; (iv) absence of confounding factors: Patients with any stage of diabetic retinopathy, including non-proliferative and proliferative forms, as well as those with diabetic macular edema, were excluded. Similarly, eyes with clinically significant epiretinal membranes, vitreomacular traction syndrome or macular holes were not included. Other exclusions comprised pathologic myopia with CNV, lacquer cracks or significant macular atrophy as well as any current or previous history of central serous chorioretinopathy and macular telangiectasia. Retinal vascular disorders, such as branch or central retinal vein occlusion and retinal artery occlusion, were also considered exclusion criteria. Furthermore, patients with inherited retinal diseases affecting macular function, such as Stargardt disease or Best disease, or those with inflammatory, infectious or degenerative chorioretinal conditions, including posterior uveitis, multifocal choroiditis or macular dystrophies, were not included in the study.

Ethics approval was granted by the Institutional Review Board of the Faculty of Medicine, LMU Munich, study ID: 20-0382, and the study followed the principles of the Declaration of Helsinki and its later amendments. Data extraction was carried out completely anonymously, with no reference to patient names or IDs. Informed consent was obtained in accordance with institutional guidelines. All patients had previously provided consent for the use of their anonymized medical data for research purposes at the time of their initial treatment or follow-up visits.

Data extraction was carried out completely anonymously, with no reference to patient names or IDs. Informed consent was obtained in accordance with institutional guidelines. All patients had previously provided consent for the use of their anonymized medical data for research purposes at the time of their initial treatment or follow-up visits. Epidemiologic data were collected for each patient, including age, gender and date of initial nAMD diagnosis.

Pre-injection examinations included assessing BCVA, measuring intraocular pressure with non-contact tonometry and performing dilated indirect fundoscopy.

Multimodal imaging was performed, including spectral-domain optical coherence tomography (OCT) and near-infrared scanning using the Spectralis HRA + OCT system (Heidelberg Engineering) at each visit. At the time of initial diagnosis, proof of CNV was primarily given by performing OCT-angiography [18, 19]. As OCT-angiography is still prone to different artifacts (e.g., projection artifacts, eye motion and motion artifacts) [20], in case of unclear OCT-angiography results, fluorescein angiography was performed to be sure about the CNV diagnosis [21].

OCT data were collected from the time of therapy initiation (mo0) and the following upload doses (mo1, mo2, mo3); patients were included when upload was finished (from mo0 to mo3) within a period of 3 ± 1 months. All OCT images included in our study underwent manual quality assessment before segmentation, ensuring that only high-quality scans were analyzed.

Biomarker segmentation was conducted using a previously trained deep learning-based semantic segmentation algorithm software from Asani et al. [17], an AI-based system designed for the automated segmentation of nAMD-related biomarkers.

Five key biomarkers of nAMD were quantitatively analyzed: CRT, IRF, SRF, SHRM and fvPED. These biomarkers were selected because of their established prognostic significance and frequent use in clinical decision-making for nAMD. They reflect both exudative changes (IRF, SRF) and fibrovascular remodeling (SHRM, fvPED).

CRT was chosen as a primary anatomical parameter, as it is a well-recognized surrogate measure for overall macular swelling and disease activity in nAMD [6]. A significant reduction in CRT often correlates with therapeutic efficacy and a decrease in exudative disease burden. IRF and SRF were analyzed separately, as both represent distinct manifestations of disease activity. IRF, located within the neurosensory retina, is strongly associated with active neovascularization and visual deterioration, whereas SRF serves as another key indicator of exudation but may have a less direct impact on visual acuity [22].

SHRM was included in the analysis as it represents hyperreflective tissue within the subretinal space, often associated with neovascular processes and having an impact on visual acuity [16]. Its reduction may indicate treatment success in suppressing pathologic neovascularization. Similarly, fvPED was quantified, as PEDs contribute to chronic exudation and their response to therapy provides further insight into disease stability and progression [6].

CRT is given in µm; volumetric biomarkers are given in arbitrary voxels as this is the more accurate way to describe the derived volumetric information (distortions caused by conversion to the metric system due to image compression can thus be reliably avoided). Given the particular interest in the development and variation of values over time, this approach appears to be the most effective method for collecting the relevant information. The algorithm provides quantitative outputs with regard to the standardized ETDRS grid [26].

The model architecture was based on a U-Net-type deep convolutional neural network [23] featuring 11 convolution layers, batch normalization [24] and ReLU activation functions [25], with transposed convolutions applied in the decoder stage to restore spatial resolution. Optimization was performed using the Adam optimizer with a categorical cross-entropy loss function to ensure effective learning. To enhance robustness and generalization, an ensemble approach was employed, wherein multiple U-Net models were trained using a leave-one-out scheme. At inference, predictions from these models were averaged to produce the final segmentation, improving consistency and mitigating potential biases in individual model predictions [17].

A total of 458 macular OCT volume scans, each obtained from a different patient, were manually annotated following the latest consensus nomenclature of the American Academy of Ophthalmology [26] by three experienced graders. Each voxel within the OCT B scans was assigned to one of the predefined classes, including IRF, SRF and fvPED, among others. To assess inter-annotator variability, an additional set of 30 scans was independently labeled by three graders.

Performance evaluation of the algorithm was conducted using the Dice-Sørensen coefficient (F1 score) [27, 28] to assess the segmentation accuracy across classes. It is derived from precision and recall, defined as the harmonic mean of these two metrics and calculated by taking twice the product of precision and recall and then dividing by their sum. It is helpful to evaluate the similarity between two sets, commonly applied in image segmentation to quantify the overlap between prediction and ground truth [29], ranging from 0 (no overlap) to 1 (perfect agreement).

On an independent test set of 36 previously unseen images, the algorithm achieved high F1 scores, with nearly perfect performance for SRF (0.98), while slightly lower scores were observed for more ambiguous features such as fvPED (0.78) and IRF (0.76) [17]. Additionally, a separate set of 30 scans was annotated by three independent graders to quantify inter-annotator variability. The algorithm’s outputs were compared against both individual annotations and the established consensus ground truth, demonstrating performance on par with human experts for most features [17].

The Shapiro-Wilk test did not indicate a normal distribution of the data. A pairwise comparison of the different biomarkers was performed using Wilcoxon matched-pairs signed rank test and Friedman-ANOVA for multiple testing if applicable. Spearman’s correlation coefficient r was used to test associations of dependent and independent variables. Errors are given as interquartile ranges (IQRs) due to non-normality of the data (Fig. 1).

Bild vergrößern

In total, 40 eyes of 38 patients were included in our analysis. Baseline demographics are summarized in Table 1. Average age at nAMD diagnosis was 78.88 ± 10.00 years (error as standard deviation); gender ratio was 17 male and 21 female. No ocular complications were recorded during this follow-up period. The distribution of MNV was as follows: 27 eyes showing type 1 MNV, 11 eyes exhibiting type 2 MNV and 2 eyes with type 3 MNV. Stratification by MNV type was not performed in the analysis, as the small sample sizes for each subtype would provide insufficient statistical power for meaningful analysis.

BCVA improved significantly within the first month of treatment, with a significant reduction in logMAR from 0.60 (IQR: 0.30) at mo0 to 0.40 (IQR: 0.40) at mo1, and this improvement was maintained through mo3 with 0.40 (IQR: 0.33). Converting the observed BCVA improvement from logMAR score to ETDRS letters, a reduction of 0.20 logMAR corresponds to a gain of approximately 10 ETDRS letters (two lines on the ETDRS chart), indicating a clinically significant improvement [30]. However, while the improvement from mo0 to mo1 was statistically and clinically significant, the difference between mo1 and mo3 (p = 0.1788) suggests a stabilization rather than a continued increase in visual acuity over the treatment period.

At baseline, all eyes exhibited MNV activity with the presence of IRF, SRF or fvPED. Biomarker changes for each visit are presented in Table 2.

At mo3, at the end of the observation period, a significant reduction of all analyzed biomarkers was present. Comparing the treatment-naïve patients (mo0) with the collective at timepoint mo3, SRF changed from 6959 (IQR: 44,545) voxel to 6 (IQR: 175) voxel, IRF from 4025 (IQR: 12,178) voxel to 19 (IQR: 94) voxel, SHRM from 1931 (IQR: 9698) voxel to 66 (IQR: 1332) voxel and fvPED from 22,045 (IQR: 59,043) voxel to 6540 (IQR: 19,316) voxel, respectively. CRT, as a surrogate marker, accordingly changed statistically significant from 433.6 (IQR: 306.6) µm to 241.5 (IQR: 130.8) µm from mo0 to mo3.

Indicative of the potent efficacy of faricimab, a significant reduction in all analyzed biomarkers was observed as early as after the first intravitreal injection at time point mo1 [SRF from 6959 (IQR: 44,545) voxel to 49 (IQR: 1221) voxel, IRF from 4025 (IQR: 12,178) voxel to 16 (IQR: 277) voxel, SHRM from 1931 (IQR: 9698) voxel to 199 (IQR: 2980) voxel and fvPED from 22,045 (IQR: 59,043) voxel to 7630 (IQR: 24,172) voxel]. Accordingly, a significant reduction of CRT [from 433.6 (IQR: 306.6) µm to 295.4 (IQR: 123.8) µm] as well as a significant improvement of BCVA [from 0.60 (IQR: 0.30) to 0.40 (IQR: 0.40)] from mo0 to mo1 was present.

Concerning the further development (mo1 to mo3), there was a significant improvement of SRF [from  49 (IQR: 1221) voxel to 6 (IQR: 175) voxel]. The other biomarkers showed no further statistically significant difference. However, the improvements of SHRM as well as fvPED were marginally not significant with p-values (mo1–mo3) of p = 0.0526 for SHRM and p = 0.0506 for fvPED, respectively. The CRT, as some kind of a collective measure for the activity parameters of nAMD, also showed a further significant change from mo1 to mo3 [from 295.4 (IQR: 123.8) µm to 241.5 (IQR: 130.8) µm].

All data as well as p values of statistical testing are shown in Table 2.

Correlations between the various biomarkers and the improvement in BCVA as an essential functional parameter were analyzed.

We found a statistically significant positive correlation of change in BCVA in logMAR (from mo0 to mo3) with CRT (r = 0.31), IRF (r = 0.32), SRF (r = 0.32) and SHRM (r = 0.41) at mo0, i.e., a high number of these activity parameters at the beginning of the intravitreal treatment provides the possibility of a huge improvement in BCVA for the patients treated. No significant correlation was found for fvPED (r = − 0.09).

Considering the correlation between change of BCVA in logMAR (from mo0 to mo3) with the change of the different biomarkers (from mo0 to mo3; marked as Δ), we found a statistically significant positive correlation with ΔIRF (r = 0.31) and ΔSHRM (r = 0.39). Weaker but still significant correlations were found for ΔSRF (r = 0.26) and ΔCRT (r = 0.22). No significant correlation was found for ΔfvPED (r = − 0.09). This suggests that IRF and SHRM are especially relevant for determining BCVA in patients with nAMD; a less pronounced correlation can be seen for BCVA and SRF as well as CRT as a summarizing parameter. Interestingly, we found no significant correlation between the change in fvPED and the change in BCVA, possibly supporting the thesis of a stable fvPED being protective against retinal atrophy and a further decline in vision [31].

Findings are summarized in Table 3.