A Systematic Review on Retinal Biomarkers to Diagnose Dementia from OCT/OCTA Images

OCT/OCTA imaging techniques and retinal layers definitions

Various imaging techniques could be leveraged to capture retinal changes. Nowadays, non-invasive imaging techniques such as OCT and OCTA are used to capture structural and vascular changes in the retina, respectively [10]. OCT is based on interference principles and can be categorized into Time domain (TD)-OCT and Frequency domain (FD)-OCT. The FD-OCT contains Spectral-Domain OCT (SD-OCT) and Swept-Source OCT (SS-OCT) [11]. FD-OCT has higher acquisition speed and better resolution replacing TD-OCT [11]. SD-OCT is commonly used in ophthalmology clinics; however, SS-OCT is a new technology with higher imaging speed and better visualization of deeper structures due to the use of longer wavelength light source [11]. SS-OCT uses more invisible light source than SD-OCT, hence, SS-OCT is relatively more comfortable for patients [11]. An example of an SD-OCT (B-scan) is shown in Fig. 2 (top). OCT can almost resolve all the retinal cellular layers as demonstrated in Fig. 2 (bottom). From the inner to outer, layers include inner limiting membrane (ILM), nerve fiber layer (NFL), ganglion cell layer (GCL), inner plexiform layer (IPL), inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL), external limiting membrane (ELM), ellipsoid zone and interdigitation zone (EZ-IZ), retinal pigment epithelium (RPE), and Bruch’s membrane (BRM).

OPEN IN VIEWER

Fig. 2

Retinal layers in a spectral domain optical coherence tomography b-scan image: unlabeled (top), labelled (bottom).

The early OCTA segmentation by [12] defined the superficial (SCP) and the deep capillary plexuses (DCP), respectively, based on the retinal anatomical layers. Simply, the SCP was defined from ILM to IPL, whereas the DCP was defined from the outer boundary of the IPL to OPL [12]. Of note, this SCP and DCP segmentation approach is commonly used by the Heidelberg SD-OCT machine [13].

Previously, in [14], the macula was analyzed into four retinal vascular networks namely: radial peripapillary capillary plexus (RPCP), superficial vascular plexus (SVP), intermediate capillary plexus (ICP), and DCP as shown in Fig. 3. The SVP was segmented between the inner 80% of the ganglion cell complex (GCC), where GCC consists of NFL+GCL+IPL. The ICP was defined as the outer 20% of the GCC to the inner 50% of the INL, whereas the DCP was defined between the outer 50% of the INL and the OPL. Recently [15], new scientific names based on segmentation boundaries were proposed where superficial vascular complex (SVC) is composed of RPCP and SVP, while deep vascular complex (DVC) is composed of ICP and DCP as shown in Fig. 3. Nevertheless, the combination of both SVC and DVC (SVC+DVC) is denoted by the inner vascular complex (IVC), as shown in Fig. 2 (bottom).

OPEN IN VIEWER

Fig. 3

Anatomic layers, previous and new optical coherence tomography angiography (OCTA) segmentation names [15]. PR, photoreceptor layer.

In order to provide a clear explanation of the various retinal layer segmentations and layer definitions in the literature, Table 1. demonstrates diverse OCT machines along with the appropriate layer definitions.

OCTA is a new emerging imaging technique that can be used to resolve retinal vasculature at the capillary level in a non-invasive manner. The En Face OCTA examples in Fig. 4 are image projections in 2D originated from a volumetric 3D OCTA around the fovea. Notably, there are various options when performing image projections at different retinal layers depths around the fovea, for instance, SVC, DVC, and IVC projections examples could be found in Fig. 4 (A1-C1). Briefly, to extract vascular parameters around the fovea, a series of steps are taken to obtain an OCTA En Face image. Firstly, a selection of retinal layers is made, for instance, from the ILM to the OPL to obtain the IVC. Next, the volumetric 3D OCTA of the selected layers is projected into a 2D image known as the En Face OCTA. This projected image will include the vessels and microcapillaries to be analyzed, providing a clear and detailed view which is ready for the extraction of vascular parameters.

OPEN IN VIEWER

Fig. 4

Examples of optical coherence tomography angiography (OCTA) scans: A1) superficial vascular complex (SVC), B1) deep vascular complex (DVC), C1) inner vascular complex (IVC), and their corresponding manual foveal avascular zone (FAZ) segmentation.

OCT/OCTA parameters

Interesting vascular changes in the retina could be observed at SVC and DVC; however, we shall report the old scientific retinal segmentation boundaries names since most of the previously published works reported the results of SCP, DCP, SVP, and ICP. These changes could be evaluated and categorized into vascular and foveal avascular zone (FAZ) changes, where manual FAZ extraction examples are shown in Fig. 4 (A2-C2). Important vascular changes could be captured via vascular density (VD) or vascular perfusion density (VPD), fractal dimension (FD), and vessel length density (VLD). Yoon and colleagues [27] defined VD/VPD as the ratio between the number of pixels in the perfused retinal vascular area to the number of pixels of the entire retina. VLD is a ratio that is defined between the total length occupied with blood vessels in an area and the total area [16]. FD is a non-linear analysis which aims to measure the geometric complexity of vessels via box counting analysis [28–30].

Generally, the process of scanning around the optic nerve head is called a disc scan mode, which could be partitioned into upper and lower sectors namely superior-hemi (S-Hemi) and inferior-hemi (I-Hemi) sectors, respectively, as shown in Fig. 5b. Furthermore, Garway-Heath et al. [31] proposed another sectioning variant in Fig. 5a, and these sections are named: superotemporal, superonasal, inferotemporal, inferonasal, nasal superior, nasal inferior, temporal superior, and temporal inferior, and denoted as ST, SN, IT, IN, NS, NI, TS, and TI, respectively. Similarly, variations around the fovea is considered as a region of interest [16, 22]; however, some papers focused on some specific regions around the fovea [21, 32–34]. As shown in Fig. 5d, a scan around the fovea could be divided into sections namely: superior (S), inferior (I), nasal (N), and temporal (T). Another simpler splitting variation around the fovea is shown in Fig. 5c with S-Hemi and I-Hemi sections.

OPEN IN VIEWER

Fig. 5

HD Angio Disc mode (optic disc) Report Layout Legend: a) Garway-Heath (GH) map [31]; b) Hemisphere Maps; 2) Angio Retina mode (macular) Report Layout Legend: c, d) Hemisphere and Quadrant Maps [32].

As illustrated in Fig. 6, the four quadrants partitions around the fovea follows a comparable pattern showed/explained previously in Fig. 5d. Additionally, SI, II, TI, and NI belongs to the inner ring; whereas SE, IE, TE, and NE belongs to the external ring around the center fovea.

OPEN IN VIEWER

Fig. 6

Microvascular network sections in superficial capillary plexuses (SCP) (adapted from [21]).

Extracting OCT/OCTA parameters methods

Some research works use the commercial software which accompanies the OCT/OCTA machine. For instance, both [19, 35] extracted OCT and OCTA parameters automatically by a commercial RTVue XR software and Optovue AngioVue system (software ReVue XR). In addition, the AngioAnalyticstrademark software was used by [36] to compute VD and FD for retinal vasculature as well as the perimeter and Acircularity Index for FAZ. Noteworthy, the Acircularity Index was introduced in [37] such that it measures the level of FAZ shape abnormalities. In the study by Chua et al. [38], Cirrus Review software generated both macular and optic disc images, whereas Iowa Reference Algorithm (IRA) [39–41] was used to automatically segment individual retinal layers. The optic disc and macular images were stitched and followed by a segmented retinal vessel map obtained by a customized built algorithm, by [38], using MATLAB software. Furthermore, Jáñez-Escalada and colleauges [42] also used IRA [39–41] as a retinal layer segmentation module (Iowa Reference Algorithms 3.6 Retinal Image Analysis Lab, Iowa Institute for Biomedical Imaging, Iowa City, IA, USA). On the contrary, some other works attempted to use deep-learning methods to develop retinal vessels as well as retinal layers segmentation methods. Based on U-Net introduced in [43], the research by [44] created a publicly available ROSE dataset and built a custom tool built specifically for the retinal vessel segmentation task of OCTA angio scans. Another work by [45] aimed to segment intraretinal layers using a cascaded version of U-Nets. In [45], the segmented layers were: ILM, mRNFL, GCL, IPL, GCIPL/GCC, INL, OPL, ONL, ELM, myoid zone, ellipsoid zone (EZ), outer segment (OS), RPE, outer layer, and Bruch’s membrane. A different study by [46] attempted to generate super-resolved segmented retinal layers using Generative Adversarial Network-based approach.

Overview of the included studies

A summary of all the longitudinal and cross-sectional (CS) studies included in this review as well as the cohort’ details, study type, imaging methods, and studied biomarkers are found in Table 2. The results are categorized into mainly structural and vascular parameters which are split cross-sectionally and longitudinally. The CS results of structural parameters include subcategories involving optic disc and macula related parameters. On the contrary, the structural and vascular parameter of the longitudinal studies are combined sequentially due to the limited number of works. Additionally, Table 3 includes the details of CS pRNFL related parameters, while Table 4 includes the longitudinal pRNFL related parameters. However, the tables for CS macula related parameters (Supplementary Table 1), CS vascular parameters (macula and optic disc) (Supplementary Table 2), and longitudinal macula-related parameters (Supplementary Table 3) are found in the Supplementary Material due to the complexity of the analyzed data. Noteworthy, all tables illustrate significant changes by ↑ or ↓ arrows, insignificant changes by - ^NS , and –for any missing parameters. Moreover, another subsection will discuss research works which attempted to build classification models based on statistically significant biomarkers.

The quality assessment results of the included research papers’ based on [9] were documented in Table 2. An example of a HQ assessed work would be the study by [53], since the diagnosis of AD was supported by Aβ₄₂, t-tau, and p-tau and the study was well organized and descriptive. On the contrary, [73] can be considered as LQ work because it failed to provide sufficient information about whether this RNFL thickness is around the optic disc or around the fovea. Nevertheless, [50] failed to distinguish dementia subtypes since patients with AD, vascular dementia (VaD), and dementia with Lewy bodies (DLB) were included in the same dementia group and hence, the quality assessment for this paper was LQ. Given that there are some discrepancies among various dementia subtypes, as discussed earlier in the review, this is a considerate drawback of the study [50]. Another study by [48] included dementia, AD, and VaD patients; however, the exact diagnosis was not clearly mentioned which is a limitation to this study. In addition, they study by [48] failed to provide the details of pRNFL thickness such that it could be an average or global pRNFL thickness, and hence inconsistence with the current review and difficult in comparison with other research works. Another limitation was observed in [34] due to not using capillaries for the vascular density of the retina, and hence, the results could potentially be improved if these small capillaries are utilized and included in the analysis.

The included studies in this review excluded all eye diseases like glaucoma and age-related macular degeneration. The only exception studies are by [33] which included pre-perimetric glaucoma (PPG) patients, as well as the study by [34] which did not include the eye health status of the cohort’s individuals.

Cross-sectional results for structural parameters

Macula-related parameters

Changes to the full retinal thickness (FRT), defined from the ILM to the top of RPE, failed to distinguish between tAD in [56], AD in [69], mild AD in [42, 47], MCI in [85], aMCI in [65], EOAD in [90], FTD in [79, 80], against HCs. Additionally, when multiple classes were involved in the analysis, the FRT variations were negligible between the three groups (mild AD, moderate AD, HCs) in [76], and between (aMCI, AD, HCs) in [67]. According to [69, 92], AD patients had a notable thinning of the GCL and IPL in contrast to HCs. In addition, in comparison to HCs, a substantial thickness reduction of GC-IPL in [25] and GCC in [17] for AD group. Moreover, both moderate AD and severe AD groups individually had a prominent thinning of GCL and IPL in [92] compared with HCs. Furthermore, a notable thickness reduction of GCL and IPL volumes was associated with AD patients unlike HCs [53]. Moreover, the outcomes of [69] indicated a significant thinning of macular RNFL (mRNFL), GCL, IPL, and ONL for AD contrasted with HCs. The results in [42] align with those in [47], indicating that a pronounced thinning of macular layers mRNFL, INL, and OSL is related to mild AD rather than to HCs. Although both works by [92] and [42] agreed that a noteworthy thinning of GCL and IPL layers is associated with mild AD unlike HCs; however, the findings of [47] is conflicting such that the thickness changes of both GCL and IPL layers were insignificant between both groups. The results reported in [17] agreed with [33] such that a notable thickness reduction of GCC layer is associated with MCI or aMCI rather than HCs. The analysis of performed by [72] indicated that the GC-IPL thickness is biomarker which is able to differentiate between three groups (MCI, dementia, HCs) as well as dementia subgroups (AD, DLB, VCID, FTD) against MCI and HCs in [72]. Additionally, changes in macula volume was significantly different among groups (AD, DLB, VCID, FTD, MCI, HCs) in [72]. When comparing the finding of [79] and [80], a majorly thinner ONL and IS-OS, also referred to as EZ, was associated with FTD and tauopathy groups in [80] as well as FTD, unknown pathology, and probable tauopathy in [79] all individually compared against HCs. Another interesting biomarker is OPL thickness changes, where the FTD groups in [79] and [80] had a significant OPL thinning compared to their corresponding HCs. The thickness of GC-IPL was substantially reduced for moderate AD and FTD with respect to HCs in [78]. Interestingly, the GC-IPL thickness variations was an excellent biomarker to detect the progression of neurodegenerative disorders in [78]. For instance, the GC-IPL thickness was considerably decreased when comparing FTD versus mild AD, moderate AD versus MCI, and moderate AD versus mild AD [78]. Other studies supported by CSF and/or brain MRI also attempted to investigate the significance of FRT changes, such that only tauopathy group had a notable FRT thinning compared with HCs; however, when adjusting the linear models, the FRT parameter was not significant anymore. The other supported studies demonstrated minor FRT changes when comparing TDP-43 and probable tauopathy groups in [79] individually against HCs.

There were other studied retinal parameters in the literature; however, some of these parameters are only associated with a single research work. For instance, significant changes to the Minimal GCL+IPL layer was an indicator to differentiate between 3 groups (MCI, Dementia, HCs) and 6 groups (MCI, AD, DLB, VCID, FTD, HCs) in [72]. In [87], the foveal macular thickness (MT), inner ring of MT in S, I, T, N quadrants, and only the T sector of the MT outer ring were notably reduced for mild AD comparatively with HCs. [76] studied the total retinal thickness (RT), total GCL thickness, RT and GCL thicknesses in S and I quadrants, as well as splitting the macula into S and I hemispheres with 5 zones each. The 5 RT zones in the S hemisphere were RT-SZ1, RT-SZ2, RT-SZ3, RT-SZ4, and RT-SZ5, whereas in the I hemispheres were RT-IZ1, RT-IZ2, RT-IZ3, RT-IZ4 and RT-IZ5. Additionally, RT-Δ1, RT-Δ2, RT-Δ3, RT-Δ4 and RT-Δ5 were all calculated by subtracting the mean inferior RT values from their corresponding mean superior RT. The previous procedure was repeated for GCL creating 5 superior zones: GCL-SZ1, GCL-SZ2, GCL-SZ3, GCL-SZ4, GCL-SZ5; 5 inferior zones: GCL-IZ1, GCL-IZ2, GCL-IZ3, GCL-IZ4, GCL-IZ5; and 5 differences (Δ): GCL-Δ1, GCL-Δ2, GCL-Δ3, GCL-Δ4, GCL-Δ5. There was no noticeable disparity between the three groups in terms of total RT and GCL thicknesses as well as RT and GCL thicknesses in S and I quadrants. Conversely, a discernible difference was found between groups, particularly in GCL-SZ2 and GCL-IZ2 thicknesses. Simply, the thickness of GCL-SZ2 was 41.26±6.95, 35.73±6.97, and 42.71±6.9 (p = 0.025) for Mild AD, Mod AD, and HCs, respectively. Moreover, the thickness of GCL-IZ2 was GCL-IZ2 39.87±6.36, 36.51±6.42, and 42.16±6.2 (p = 0.048) for Mild AD, Mod AD, and HCs, respectively. The rest of the parameters were insignificant between the groups. Garway-Heath (GH) map introduced in [31], shown in Fig. 5, was implemented by [32] to include two splitting patterns (quadrants) around the optic disc as well as around the fovea. The OCT optic disc parameters were comprised of the overall pRNFL thickness and pRNFL following GH in S & I Hemisphere (Hemi) and GH in 8 sections illustrated in Fig. 5.-a. Besides, the OCTA optic disc parameters were the VD of RPC (whole and peripapillary region), RPC VD pursuing GH in S & I Hemi and GH in 8 sections. On the other hand, the OCT foveal parameters were whole foveal thickness, macular thickness four quadrants (S, I, T, N), and macular thickness adopting GH in S & I Hemi method. Additionally, the OCTA foveal parameters contained the VD of SCP and DCP each evaluated at whole fovea, total parafovea, parafovea S & I hemi method, and parafovea in four quadrants. Only TI, IT, TS, and ST GH sections of pRNFL thicknesses were insignificant between the groups; however, all other OCT optic disc parameters were significantly thinner for AD group compared to HCs (p < 0.05). Additionally, all OCTA optic disc parameters (except VD of RPC GH-IT p = 0.066) were statistically significant between the groups. Oppositely, the OCT macular thickness parameters failed to distinguish between the groups (p > 0.05); however, the OCTA VD of macular SCP (I-Hemi: p = 0.026, T: p = 0.007, N: p = 0.003), and DCP (I-Hemi: p = 0.049, T: p = 0.013) were significantly reduced in mild AD patients compared to HCs. These OCT/OCTA significant changes indicates a damage to both the retinal microvascular system and retina neurons which is linked with cognitive decline. [32] correlated VD of the retina with memory, overall cognition, and visual-spatial perception Functions. The SN sector of peripapillary RPC VD can predict overall cognitive, executive, and visuospatial functions. Additionally, the NS and TS sectors of peripapillary RPC VD can predict memory function. Furthermore, the NI sector of peripapillary RPC VD can predict visuospatial function. Moreover, the T and S-Hemi sectors of macular SCP VD can predict executive function, while the T sectors of macular DCP VD can predict executive function. The reduced retinal VD is associated to AD; however, the novelty of [32] is the particularity in showcasing exactly which areas of the retina are VD reduced.

Cross-sectional results for vascular parameters

Longitudinal results for both structural and vascular parameters

Interestingly, the total pRNFL thinning rate for AD patients (over 12 months) was –1.6±2.3μm which is greater than the –0.6±1.4μm (p-value=0.04) in [51] and the average of 0.5μm associated with natural thinning due to aging [95]. The pRNFL thinning rate in the S section was –2.1±3.0 and –0.5±3.3 for AD and HCs (p-value=0.01), respectively. Moreover, the pRNFL thinning rate in the I section was –2.9±3.2 for AD patients which is far greater than the thinning rate of –0.6±2.5 for HCs (p-value=0.001), respectively. Conversely, the thinning rate of pRNFL in T and N quadrants were negligible. Hence, [51] concludes that a higher thinning rate of pRNFL (S, I, total) over longitudinal visits may indicate cognitive decline. Another longitudinal study by [70] which analyzed pRNFL thickness changes over the course of 25 month. The thickness of global pRNFL as well as the thicknesses in S, I, T, and N quadrants were assessed for 20 MCI patients and 58 HCs initially. Later in the study, 9 MCI patients converted to Mild-AD, 1 MCI converted to Moderate-AD, and 10 MCI patients remained stable. Additionally, 8 HCs converted to MCI and 50 HCs remained stable, and hence, there is a total of 60 and 18 stable and converted participants, respectively. This study focused on comparing stable and converted groups and only the pRNFL thickness in the I sector was significantly reduced between both groups. In fact, the converted group had an pRNFL I thickness of –11.0±12.8 μm whereas the stable group had a thickness of 0.4±15.7 μm (p = 0.009). The results by [70] imply that patients with a higher thickness reduction of pRNFL in the I sector could potentially be more susceptible to switching to a worse cognitive state, such as from MCI to AD dementia.

In terms of structural layer thickness, a longitudinal study by [83] evaluated the GCL, IPL, INL, OPL, ONL, mRNFL thicknesses (outer, inner, total), as well as the average of pRNFL thicknesses for 430 participants initially. Later, the study cohort were split into Drop-out and Follow-up groups each 215 individuals. The MMSE scores decreased more rapidly throughout the follow-up period when the total mRNFL thickness at baseline was thinner (p = 0.01). When compared to the dropout group, the follow-up group had a moderately thicker retinal layer specifically in the inner mRNFL (p = 0.001), total mRNFL (p = 0.04), outer GCL outer (p = 0.02), and inner INL (p = 0.04) at the baseline evaluation.

A stronger decline in the MMSE scores over the follow-up period was correlated with an initial thickness reduction of the total mRNFL at baseline (p = 0.01). Moreover, a baseline attenuated outer and inner thicknesses was linked with a decrease in MMSE scores with p = 0.04 and p = 0.01, respectively. [83] concluded that individuals with a reduced baseline mRNFL were more susceptible to suffer from cognitive impairment compared to their counterparts.

Another longitudinal study by [52] explored the thicknesses of pRNFL (Avg, S, I, T, N), GC-IPL (mean, min, ST, S, SN, IN, I, IT), central macular retinal thickness (MRT), mean macular cube thickness, as well as the macular cube volume for a cohort of 42 AD, 26 aMCI, and 66 HCs. [52] study was keen on obtaining neuropsychological tests for MCI and AD patients once every year for two years through follow ups, hence this study had access to some MCI patients who became AD and were represented by MCI _AD . The other group of MCI patients who remained with the same diagnosis, or simply with a stabled cognitive state, were denoted by (MCI _MCI ). Unlike the 12 MCI _MCI and based on this study, nine MCI _AD had a significant reduction in the T section of pRNFL thickness, the average and minimum GC-IPL thicknesses, and the GC-IPL thickness in the IN, I, and IT sectors. Hence, [52] proved that macular GC-IPL parameters are better in contrast to pRNFL parameters especially in predicting the progression of MCI to AD-type dementia. Conversely, [52] excluded patients with glaucoma, and hence the association between GC-IPL thinning and glaucoma is still unraveled. Simply, GC-IPL thinning might also indicate glaucoma along with AD.

A different longitudinal study by [69] segmented retinal layers by two methods namely N-site Axonal software by Spectralis and a new segmentation technology of the Spectralis. The N-site Axonal software analyzed only the initial cohort of 150 AD, whereas the new segmentation method assessed the retinal layer for follow-ups. They study by [69] initially involved 150 AD patients and 75 age matched HCs; however, only 99 and 51 AD patients were successful in follow-ups<3 years and≥3 years, respectively. The N-site software provided the average pRNFL thickness, the pRNFL thickness ratio between N/T, as well as the thicknesses of papillomacular bundle (PapMac-Bundle) and in quadrants: NS, N, NI, TI, T, and TS. Compared to HCs, the Avg pRNFL, and RNFL in TI, T, TS, PapMac-Bundle, and pRNFL N/T ratio were significantly lower for AD group. On the other hand, the new segmentation method provided the thickness of specific layers ILM, RNFL, GCL, IPL, INL, OPL, ONL, OLM, Photoreceptors, and RPE for various comparisons including: AD versus HCs, AD < 3 years versus AD≥3 years, and HCs versus AD < 3 years versus AD≥3 years. When comparing AD versus HCs, the thicknesses of mRNFL, GCL, IPL, and ONL were notably reduced for AD group conversely to HCs. Surprisingly, both comparisons between AD < 3 years versus AD≥3 as well as the three groups (HCs versus AD < 3 years versus AD≥3 years) were significantly different in mRNFL, GCL, and IPL thicknesses, which indicates that these parameters could prediction the progression of AD.

A longitudinal-based study by [60] focused mainly on IVC FAZ area as well as the annual change of FAZ area between both groups, and these parameters were recorded over the course of 3 consecutive years. Enlarged IVC FAZ was statistically significant (p = 0.031) and was associated with BIOM_+ve group. Specifically, the FAZ in mm² were 0.272±0.083 and 0.368 ± 0.077 for BIOM_-ve and BIOM_+ve, respectively. However, there was no statistically significant change seen at follow-up (p > 0.05) for the annual FAZ parameter. A longitudinal research for the course of 2 years in [35] studied OCT parameters (pRNFL, GCC thicknesses) and OCTA parameters (FAZ area, VD of SCP, DCP, CC, and RPC) changes at baseline and after follow-up for aMCI and HCs. The cohort of 18 HCs and 19 aMCI patients were split into 12 stable aMCI (aMCI _Stable ) patients and 7 aMCI patients converted to (AD-type) dementia (aMCI _Converted ). Interestingly, the thickness of GCC and pRNFL were significantly reduced for aMCI patients over follow-ups, while no significant changes were noticed for HCs. Additionally, the VD of SCP, DCP, RPC plexus were reduced while FAZ area was increased for aMCI at follow-up in contrast to baseline. Conversely, the VD of CC reduced but did not reach statistical significance between aMCI at baseline and at follow-up. Furthermore, [35] performed a comparison between aMCI _Converted and aMCI _Stable in terms of altered OCT/OCTA parameters. They found out that GCC and pRNFL thicknesses were significantly reduced for both aMCI _Converted and aMCI _Stable sub-categories. In addition, the VD of SCP, DCP, and RPC were reduced, while the FAZ area was increased for aMCI _Stable group at follow-up. On the other hand, only the VD of SCP and RPC where significantly reduced, while the FAZ area was increased for aMCI _Converted group at follow-up. However, there were no significant alterations to the VD of CC in both aMCI _Converted  and aMCI _Stable  sub-groups, while VD of DCP was also insignificantly changed for only aMCI _Converted patients.

Diagnosis-related research works

Some research articles attempted to evaluate the diagnostic values for retinal biomarkers in terms of classifying neurodegenerative disorders. Starting with a longitudinal study by [59] which took the course of 25 months. The initial cohort in [59] consisted of 20 MCI and 58 HCs, and the follow-up resulted in 8 HCs converted to MCI, 9 MCI converted to mild-AD, and one MCI patient converted to moderate-AD. [59] combined the stable individuals (from both HCs and MCI groups) to form 60 Stable _ Total group and combined the converted individuals to form 18 Converted _ Total. Using baseline story recall as a predictor, the multivariable logistic regression model’s AUC was 0.854; however, when the longitudinal reduction in pRNFL thickness in the inferior quadrant was included, the AUC improved to 0.915. In order to calculate the likelihood of cognitive decline, [59] created a formula based on the pRNFL episodic memory model, and this method produced a conversion score. In comparison to the stable participants, the converted participants’ conversion score was considerably higher (0.59±0.30 versus 0.12±0.19, p < 0.001). [59] picked the best conversion score cut-off value of 0.134 which produced a sensitivity and a specificity of 0.944 and 0.767, respectively, in identifying cognitive decline.

Noteworthy, the remaining diagnostic-based studies in this review, [26, 92], involved cross-sectional cohorts only. According to the findings of a study [72], the thickness of the GCL+IPL is the most informative indicator of brain health condition. The study also suggested that a thickness cutoff of 75μm yielded the best balance between sensitivity (85%) and specificity (61%) in a sample comprising 66 individuals with MCI, 27 HCs, and 43 individuals with various types of dementia (including 17 cases of AD, 16 cases of DLB, 6 cases of VCID, and 4 cases of FTD). Individuals with GCL+IPL thickness less than 75μm have lower Montreal Cognitive Assessment scores after Bonferroni correction for multiple comparisons. [72] used logistic regression analysis and found that GCL+IPL thickness had the highest discriminability in distinguishing between healthy controls and individuals with CI, with an AUC of 0.821. The minimum GCL+IPL thickness had the second-highest discriminability, with 0.812 AUC.

In [57], the full retinal thickness (FRT) changes were insignificant between the groups (AD, PD, HCs); however, the issue was mainly caused by mRNFL layer which was then was rolled out due to being a common feature in multiple diseases. Therefore, [57] focused on analyzing texture characteristics of fundus images obtained from OCT scans, and used support vector machines to build a classification model for the purpose of differentiating between the three classes. When building the classification models, removing mRNFL layer achieved better classification results with sensitivity, specificity, and accuracy of 87.0, 100.0, and 82.9, respectively. Participants received the same classification for both eyes with median accuracy percentages of 91.4 (2-fold cross-validation) up to 94.4 (10-fold cross-validation). [57] proved that retinal thickness could be utilized to classify and distinguish between AD, PD, and HCs. The research done by [75] attempted to use logistic regression to distinguish between CH-PAT and CH-NAT groups using the pRNFL thickness in S, I, T, N quadrants as predictors. Their analysis indicated that OD_T and OD_N (OD = right eye) combined predictors resulted in a model with 87% sensitivity and 56% specificity (AUC = 0.83).

Redirecting our attention to research works by [34, 38] that studied both structural and vascular parameters. In [34], the receiver operating characteristics (ROC) curves demonstrated sensitivities (represented as true-positive rate) and specificities (also known as false-positive rate) for various parameters. Also in [34], the area under the ROC (AUC) based on mRNFL+GCL thickness was 0.874, while based on VD parameters was the AUC was 0.696 in terms of differentiating between AD from HCs. Notably, when combining mRNFL+GCL thicknesses as well as the vascular parameters, the ROC curve improved to 0.892 in diagnosing AD patients, and hence better sensitivity. On the contrary, [38] aimed to distinguish between MCI and/or AD with HCs by evaluating the diagnostic performance of retinal VD, measured or compensated cpRNFL, and macular layers (mRNFL, mGCL, and mIPL) parameters. When comparing all groups together, or MCI against HCs, or AD against HCs, or MCI against AD, the retinal VD, and macular layers parameters all failed to draw a line between the groups (p > 0.05). Additionally, the compensated cpRNFL thickness also failed to differentiate between MCI and AD groups. However, the compensated cpRNFL thickness parameter surpassed the measured cpRNFL in discriminating between the 3 groups (AUC = 0.74 versus 0.69 with p = 0.026), between MCI against HCs (AUC = 0.74 versus 0.68 with p = 0.020), and AD against HCs (AUC = 0.79 versus 0.71 with p = 0.025). Furthermore, when combining mGCC with either compensated (Model#3) or measured cpRNFL (Model#4), a classification improvement was achieved, compared to the measured cpRNFL, for the 3 groups (AUC = 0.80 versus 0.69 with p < 0.001), between MCI against HCs (AUC = 0.79 versus 0.68 with p < 0.001), AD against HCs (AUC = 0.87 versus 0.71 with p < 0.001), and MCI against AD (AUC = 0.72 versus 0.58 with p = 0.003). Finally, mGCC outperformed mGC-IPL to separate MCI and HCs groups (AUC = 0.71 versus 0.66 with p = 0.038); however, mGCC and mGCIP failed to separate AD from HCs due to statistical insignificance.

Interestingly, both [26, 92] research works examined six popular machine learning (ML) techniques, including extreme gradient boosting, light gradient boosting machine, k-nearest neighbor, random forest, gradient boosting, and adaptive boosting, were examined on slightly different cohorts. The accuracy, AUC, f1 score, and recall were the ML assessment metrics that were used. In [26], all six popular ML techniques, were examined on the vascular parameters of 77 AD and 145 HCs cohort. The diagnosis models were developed to include OCTA confounders with statistically significant p-values. The cohort were randomly split into training and testing dataset with a ratio of 7 : 3, where the adaptive boosting algorithm demonstrated the highest diagnostic performance for the testing set with 0.75, 0.73, 0.72, and 0.75 for accuracy, AUC, f1 score, and recall, respectively. On the contrary, in [92], the structural parameters of a cohort composed of 159 AD and 299 HCs were assessed by the same six popular ML techniques used in [26]. Similarly to [26], a random split with 7 : 3 ratio was performed to create training and testing datasets, respectively. The XGBoost algorithm demonstrated the most effective diagnostic performance, with testing scores for the accuracy, AUC, f1 score, and recall of 0.74, 0.69, 0.70, and 0.74, respectively. Except for the Light GBM model, the MRT confounder showed utter superiority in the five ML models. Conversely, the mRNFL thickness was the crucial parameter for the Light GBM model.

Concluding remarks

The inclusion and exclusion criteria for the review, illustrated in Fig. 1, led to the incorporation of 64 original papers included in this investigation. The scope of this study involves non-invasive imaging techniques OCT & OCTA and their potential in diagnosing dementia disorders. Furthermore, the most prevalent dementia diseases, including AD, VaD, DLB, FTD, MixDem, and a few PD-based papers with AD/MCI, were simply described in this paper.

The results of the current work indicate that the pRNFL-S is a trustworthy biomarker to detect most AD cases; however, its effectiveness may be impacted when using a C-SD OCT machine, and hence, additional research is needed in future studies. In addition, the pRNFL-S biomarker showed no discriminative power between mild AD and HCs in [87, 92]. Similarly, the dependability of pRNFL-S as a biomarker for detecting MCI is also restricted by specific OCT machines, such that the optimal effectiveness is achieved when using A-SD OCT machines. Furthermore, recent studies confirmed the capability of pRNFL-I biomarker in recognizing AD is limited to only A-SD [17, 64], while deemed unreliable when other OCT machines are used.

Another reliable biomarker, the pRNFL-G, effectively differentiates between FTD from mild AD and HCs in [78], moderate AD from HCs in [78], MCI from HCs in [53, 78], as well as detecting pathological Aβ₄₂/tau in cognitively healthy individuals [75]. Yet, the pitfall of the pRNFL-G biomarker is the contradictory results in findings by [78] and [82]. Simply, a significant thinning of pRNFL-G was linked with mild AD in [82], while negligible pRNFL-G thickness changes were documented in [78] between mild AD and HCs. Another limitation of pRNFL-G is its inability to recognize the progression of neurodegenerative disorders [78], and hence was also not capable to discriminate between multiple groups with varied degrees of cognitive impairment [50, 66].

In addition to the previously discussed biomarkers, emerging evidence suggests that the pRNFL-Avg thinning is considered as viable biomarker to discriminate between AD and HCs, except when using C-SD and O-SD OCT machines which should be further investigated in future works. Additionally, the pRNFL-Avg was capable of discriminating between severe and mild AD [92], and hence, assessing the worsening stage of AD. Moreover, pRNFL-Avg thinning could be regarded as a dependable biomarker to differentiate between MCI and HCs; however, its effectiveness is questionable when using DRI SS-OCT Triton-Topcon (T-SS) or C-HD OCT machines and when the size of the cohort is relatively low.

It is also worth noting that recent studies by [20, 88] proved the robustness of choroidal thickness as a retinal biomarker to differentiate between AD and HCs groups. On the contrary, the utilization of choroidal thickness should be limited to non-AMD patients, because choroidal thinning is a trait shared by both AMD and AD in the elderly [96].

Pivoting our focus to another category of biomarkers, vasculature-based changes that are linked with cognitive impairment include FAZ enlargement, as well as alterations around FAZ, VD reduction in ICP and DCP (VPD and VLD) for cognitively impaired individuals in contrast to HCs. However, there were no sufficient evidence to support this correlation, as only a few works reported significant vasculature variations while other works were contradicting. Additionally, the process of comparing the vascular-based parameters was challenging due to the various layers’ definitions adopted by researchers, where a summary was previously shown in Table 1. With regards to the enlargement of FAZ area, a few things must be considered. Firstly, the retinal layers involved prior retrieving En-Face OCTA image projection, for instance, the FAZ of IVC will be much different than SCP. Secondly, the FAZ parameters extracted by various OCT machines could be different due to the hardware operation differences of these machines or related to the segmentation methods used to extract FAZ area. Thirdly, FAZ parameters are affected by other factors like myopic eyes, and these parameters could be adjusted with axial length information [97]. Bearing in mind the previous limitations, the FAZ area changes has limited capabilities to detect cognitive impairment.

A general problem was identified by the review which is the unavailability of certain parameters. Unlike the pRNFL thickness, the process of determining the appropriate retinal layers to identify neurodegenerative disorders is quite challenging due to the discrepancy between the findings. Simply, different works studied some parameters which are not found in other research articles. For instance, GC-IPL thickness was studied in [98] but not in all research articles in the literature.

The diagnostic capabilities of various retinal biomarkers have been explored in this review. Most of these biomarkers are based on structural thicknesses, while other biomarkers are based on vascular changes. On the other hand, we hypothesize that combining multiple OCT/OCTA biomarkers will enhance the overall diagnostic accuracy for the task of recognizing distinct neurodegenerative disorders. Our assumptions were supported by [34], where the incorporation of both structural thicknesses and vascular parameters yielded in a better AUC results, and consequently, this would potentially be the future direction of this field of study.

The potential limitation of this review could be the fact of not involving pre-clinical AD and focusing on other works related to dementia. Based on the discussed methods of extracting retinal structural and/or vascular parameters, there are a few points to be considered. In short, the segmentation of both retinal vasculature and layers is a crucial step before extracting retinal parameters. Errors in segmentation may lead to incorrect extracted parameters and hence the retinal analysis may fail to realize significant outcomes or the outcome could be misleading. Researchers tend to perform manual corrections to the segmented images; however, this process is not only time-consuming but also requires great expertise in the medical field which potentially not feasible. Therefore, next future work should focus on developing and improving segmentation approaches to efficiently/accurately automate the procedure of retinal layers extraction. Noteworthy, improving the segmentation efficacy would open the door to investigate more biomarkers that initially thought to be irrelevant to various disorders.